a novel methodology for planning reliable wireless sensor...

UNIVERSIDAD POLITÉCNICA DE MADRID

DEPARTAMENTO DE AUTOMÁTICA, INGENIERÍA ELECTRÓNICA E INFORMATICA INDUSTRIAL

ESCUELA TÉCNICA SUPERIOR DE INGENIEROS INDUSTRIALES

CENTRO DE ELECTRÓNICA INDUSTRIAL

A novel methodology for planning reliable wireless sensor networks

TESIS DOCTORAL

Autor: Danping He Master of Electronics Engineering from Politecnico di Torino

Directores: Teresa Riesgo Alcaide Doctora Ingeniera Industrial por la Universidad Politécnica de Madrid Jorge Portilla Berrueco Doctor por la Universidad Politécnica de Madrid en Ingeniería Electrónica

2014

Tribunal

Tribunal nombrado por el Mgfco. y Excmo. Sr. Rector de la Universidad Politécnica de Madrid, el día de de 2013. Presidente: Javier Uceda, Universidad Politécnica de Madrid

Vocales: Roberto Sarmiento Rodríguez, Universidad de las Palmas de Gran Canaria

Celia López Ongil, Universidad Carlos III de Madrid

Alan Mc Gibney, Cork Institute of Technology

Secretario: José Ramón Casar, Universidad Politécnica de Madrid

Suplentes: Ángel De Castro Martín, Universidad Autónoma de Madrid

Marta Portela García, Universidad Carlos III de Madrid

Realizado el acto de lectura y defensa de la Tesis el día de de 2013 en la Escuela Técnica Superior de Ingenieros Industriales de la Universidad Politécnica de Madrid. Calificación: EL PRESIDENTE LOS VOCALES EL SECRETARIO

�

To my parents

Acknowledgements

I came to this beautiful country three years ago to pursue PhD study in CEI. I’ve

been getting acquainted with those lovely professors and colleagues, who inspire me

the best on scientific researches and bring a lot of happiness to my life. I do not feel

alone on the way fighting for dreams, because deep in my heart, I know that there are

someone always besides me and supporting me. At this important moment when I

about to finish this doctoral thesis, I would like to express my sincerely appreciations

to them.

First and foremost, I would like to thank my advisor Professor Teresa Riesgo. I’ve

been profoundly benefit from her advises and inspirations over the past three years,

which are always important guiding lights leading me to the bright way towards

scientific research. I deeply thank her for the unprecedented freedom she offered to

explore my intellectual curiosity in my work, and for fostering my capacity critically

as an independent researcher.

I would also sincerely thank my co-advisor Dr. Jorge Portilla. I absorbed

important knowledge from him through every discussion, not only on professional

area but also on other aspects that will absolutely benefit my future careers. I learnt

a lot from him and progressed fast. I will always appreciate him for all the patience

to correct my work, and for the consistent support to eliminate my nervousness

before presentations.

I take this opportunity to record my gratitude to Professor David Symplot-Ryl

and Dr. Nathalie Mitton, who offered me the exchange research stay at Inria

Lille - Nord Europe. Their valuable advices and ideas not only make me

successful in academic publications, but also illuminate another research direction

and methodology to me.

I would like to thank the members of my thesis committee, for generously offering

their time, support, guidance and good will throughout the preparation and review

of this document.

Gabriel Mujica, my project partner in CEI. We work together for the great

aspiration of achieving success in DPCM project. His knowledge on hardware

programming perfectly compensates my shortage and it’s my fortune to have his

collaboration. I would like to thank him for all the intelligent and patient works

on establishing sensor testbed to allow us making real measurements efficiently and

precisely.

I would also express my acknowledgement to Edu, who gives me so many

important advises and encouragements. From time to time, he concerns about my

work and my progress on preparing documents for the thesis committee. Without

his great efforts, the whole procedure would not be run smoothly.(��!)

All the lovely professors in CEI, Jose Antonio, Yago, Rafa, Jesus, Pedro and

Roberto. I sincerely thank you all for the supports and encouragements on my

study. I gain quite significantly from your teaching and supports.

I enjoyed a wonderful time in CEI with all the colleagues (CEIers). They are

earnest, friendly and hard-working, and I would like to appreciate them for all the

help and happiness they gave me. My special gratitude should give to Marcelo who

taught me on how to efficiently use Latex, Monica who translated for me the abstract

of this thesis, and Nico who gave me so many helpful suggestions on multi-objective

optimization and writing papers, thank you so much!

I would like to express my deep appreciation to James Zhao for everything he

has done for me. His working attitude inspire me so much along these years. His

valuable advises, encouragements and generous supports bring warmth to me and

shine my life, like the sunlight in severe winter.

My acknowledgement will never be complete without the special mention of

my Chinese friends. I would like to thank Tianjun Zhou, Cheng Xie, Yang Wang,

Pengming Cheng, Wei He, Sisi Zhao, Zhi Wang and Ke Guan for being with me and

trusting me. Their moral supports and motivations drive me to do the best and I

find myself lucky to have friends like them.

Finally, I would like to give my greatest appreciation to my parents Dedong He

and Sujuan Qiu, for showing faith in me and giving me liberty to chase what I

desired, for standing behind me with all their love.

Sincerely thank you all. Danping

��

��

ii

Abstract

Wireless sensor networks (WSNs) have shown their potentials in various

applications, which bring a lot of benefits to users from both research and industrial

areas. For many setups, it is envisioned that WSNs will consist of tens to hundreds of

nodes that operate on small batteries. However due to the diversity of the deployed

environments and resource constraints on radio communication, sensing ability and

energy supply, it is a very challenging issue to plan optimized WSN topology and

predict its performance before real deployment.

During the network planning phase, the connectivity, coverage, cost, network

longevity and service quality should all be considered. Therefore it requires designers

coping with comprehensive and interdisciplinary knowledge, including networking,

radio engineering, embedded system and so on, in order to efficiently construct

a reliable WSN for any specific types of environment. Nowadays there is still a

lack of the analysis and experiences to guide WSN designers to efficiently construct

WSN topology successfully without many trials. Therefore, simulation is a feasible

approach to the quantitative analysis of the performance of wireless sensor networks.

However the existing planning algorithms and tools, to some extent, have serious

limitations to practically design reliable WSN topology:

• Only a few of them tackle the 3D deployment issue, and an overwhelming

number of works are proposed to place devices in 2D scheme. Without

considering the full dimension, the impacts of environment to the performance

of WSN are not completely studied, thus the values of evaluated metrics such

as connectivity and sensing coverage are not sufficiently accurate to make

proper decision.

• Even fewer planning methods model the sensing coverage and radio

propagation by considering the realistic scenario where obstacles exist. Radio

signals propagate with multi-path phenomenon in the real world, in which

direct paths, reflected paths and diffracted paths contribute to the received

signal strength. Besides, obstacles between the path of sensor and objects

might block the sensing signals, thus create coverage hole in the application.

• None of the existing planning algorithms model the network longevity and

packet delivery capability properly and practically. They often employ

unilateral and unrealistic formulations.

• The optimization targets are often one-sided in the current works. Without

comprehensive evaluation on the important metrics, the performance of

planned WSNs can not be reliable and entirely optimized.

• Modeling of environment is usually time consuming and the cost is very

high, while none of the current works figure out any method to model

the 3D deployment environment efficiently and accurately. Therefore many

researchers are trapped by this issue, and their algorithms can only be

evaluated in the same scenario, without the possibility to test the robustness

and feasibility for implementations in different environments.

In this thesis, we propose a novel planning methodology and an

intelligent WSN planning tool to assist WSN designers efficiently

planning reliable WSNs.

First of all, a new method is proposed to efficiently and automatically

model the 3D indoor and outdoor environments. To the best of our

knowledge, this is the first time that the advantages of image understanding

algorithm are applied to automatically reconstruct 3D outdoor and indoor scenarios

for signal propagation and network planning purpose. The experimental results

indicate that the proposed methodology is able to accurately recognize different

objects from the satellite images of the outdoor target regions and from the

scanned floor plan of indoor area. Its mechanism offers users a flexibility to

reconstruct different types of environment without any human interaction. Thereby

it significantly reduces human efforts, cost and time spent on reconstructing a 3D

geographic database and allows WSN designers concentrating on the planning issues.

Secondly, an efficient ray-tracing engine is developed to accurately

and practically model the radio propagation and sensing signal on the

constructed 3D map. The engine contributes on efficiency and accuracy to the

estimated results. By using image processing concepts, including the kd-tree space

division algorithm and modified polar sweep algorithm, the rays are traced efficiently

without detecting all the primitives in the scene. The radio propagation model

iv

is proposed, which emphasizes not only the materials of obstacles but also their

locations along the signal path. The sensing signal of sensor nodes, which is sensitive

to the obstacles, is benefit from the ray-tracing algorithm via obstacle detection.

The performance of this modelling method is robust and accurate compared with

conventional methods, and experimental results imply that this methodology is

suitable for both outdoor urban scenes and indoor environments. Moreover, it can

be applied to either GSM communication or ZigBee protocol by varying frequency

parameter of the radio propagation model.

Thirdly, WSN planning method is proposed to tackle the above

mentioned challenges and efficiently deploy reliable WSNs. More metrics

(connectivity, coverage, cost, lifetime, packet latency and packet drop rate) are

modeled more practically compared with other works. Especially 3D ray tracing

method is used to model the radio link and sensing signal which are sensitive to the

obstruction of obstacles; network routing is constructed by using AODV protocol;

the network longevity, packet delay and packet drop rate are obtained via simulating

practical events in WSNet simulator, which to the best of our knowledge, is the

first time that network simulator is involved in a planning algorithm. Moreover, a

multi-objective optimization algorithm is developed to cater for the characteristics

of WSNs. The capability of providing multiple optimized solutions simultaneously

allows users making their own decisions accordingly, and the results are more

comprehensively optimized compared with other state-of-the-art algorithms.

iMOST is developed by integrating the introduced algorithms, to

assist WSN designers efficiently planning reliable WSNs for different

configurations. The abbreviated name iMOST stands for an Intelligent

Multi-objective Optimization Sensor network planning Tool. iMOST contributes

on: (1) Convenient operation with a user-friendly vision system; (2) Efficient and

automatic 3D database reconstruction and fast 3D objects design for both indoor

and outdoor environments; (3) It provides multiple multi-objective optimized 3D

deployment solutions and allows users to configure the network properties, hence it

can adapt to various WSN applications; (4) Deployment solutions in the 3D space

and the corresponding evaluated performance are visually presented to users; and (5)

The Node Placement Module of iMOST is available online as well as the source code

of the other two rebuilt heuristics. Therefore WSN designers will be benefit from

v

this tool on efficiently constructing environment database, practically and efficiently

planning reliable WSNs for both outdoor and indoor applications. With the open

source codes, they are also able to compare their developed algorithms with ours to

contribute to this academic field.

Finally, solid real results are obtained for both indoor and outdoor

WSN planning. Deployments have been realized for both indoor and outdoor

environments based on the provided planning solutions. The measured results

coincide well with the estimated results. The proposed planning algorithm is

adaptable according to the WSN designer’s desirability and configuration, and it

offers flexibility to plan small and large scale, indoor and outdoor 3D deployments.

The thesis is organized in 7 chapters. In Chapter 1, WSN applications and

motivations of this work are introduced, the state-of-the-art planning algorithms and

tools are reviewed, challenges are stated out and the proposed methodology is briefly

introduced. In Chapter 2, the proposed 3D environment reconstruction methodology

is introduced and its performance is evaluated for both outdoor and indoor

environment. The developed ray-tracing engine and proposed radio propagation

modelling method are described in details in Chapter 3, their performances are

evaluated in terms of computation efficiency and accuracy. Chapter 4 presents

the modelling of important metrics of WSNs and the proposed multi-objective

optimization planning algorithm, the performance is compared with the other

state-of-the-art planning algorithms. The intelligent WSN planning tool iMOST is

described in Chapter 5. Real WSN deployments are prosecuted based on the planned

solutions for both indoor and outdoor scenarios, important data are measured and

results are analysed in Chapter 6. Chapter 7 concludes the thesis and discusses

about future works.

vi

Resumen en Castellano

Las redes de sensores inalambricas (en ingles Wireless Sensor Networks, WSNs) han

demostrado su potencial en diversas aplicaciones que aportan una gran cantidad

de beneficios para el campo de la investigacion y de la industria. Para muchas

configuraciones se preve que las WSNs consistiran en decenas o cientos de nodos

que funcionaran con baterıas pequenas. Sin embargo, debido a la diversidad de

los ambientes para desplegar las redes y a las limitaciones de recursos en materia

de comunicacion de radio, capacidad de deteccion y suministro de energıa, la

planificacion de la topologıa de la red y la prediccion de su rendimiento es un tema

muy difıcil de tratar antes de la implementacion real.

Durante la fase de planificacion del despliegue de la red se deben considerar

aspectos como la conectividad, la cobertura, el coste, la longevidad de la red y

la calidad del servicio. Por lo tanto, requiere de disenadores con un amplio e

interdisciplinario nivel de conocimiento que incluye la creacion de redes, la ingenierıa

de radio y los sistemas embebidos entre otros, con el fin de construir de manera

eficiente una WSN confiable para cualquier tipo de entorno. Hoy en dıa todavıa

hay una falta de analisis y experiencias que orienten a los disenadores de WSN

para construir las topologıas WSN de manera eficiente sin realizar muchas pruebas.

Por lo tanto, la simulacion es un enfoque viable para el analisis cuantitativo del

rendimiento de las redes de sensores inalambricos.

Sin embargo, los algoritmos y herramientas de planificacion existentes tienen, en

cierta medida, serias limitaciones para disenar en la practica una topologıa fiable de

WSN:

• Solo unos pocos abordan la cuestion del despliegue 3D mientras que existe una

gran cantidad de trabajos que colocan los dispositivos en 2D. Si no se analiza la

dimension completa (3D), los efectos del entorno en el desempeno de WSN no

se estudian por completo, por lo que los valores de los parametros evaluados,

como la conectividad y la cobertura de deteccion, no son lo suficientemente

precisos para tomar la decision correcta.

• Aun en menor medida los metodos de planificacion modelan la cobertura de

los sensores y la propagacion de la senal de radio teniendo en cuenta un

escenario realista donde existan obstaculos. Las senales de radio en el mundo

real siguen una propagacion multicamino, en la que los caminos directos, los

caminos reflejados y los caminos difractados contribuyen a la intensidad de

la senal recibida. Ademas, los obstaculos entre el recorrido del sensor y los

objetos pueden bloquear las senales de deteccion y por lo tanto crear areas sin

cobertura en la aplicacion.

• Ninguno de los algoritmos de planificacion existentes modelan el tiempo

de vida de la red y la capacidad de entrega de paquetes correctamente

y practicamente. A menudo se emplean formulaciones unilaterales y poco

realistas.

• Los objetivos de optimizacion son a menudo tratados unilateralmente en

los trabajos actuales. Sin una evaluacion exhaustiva de los parametros

importantes, el rendimiento previsto de las redes inalambricas de sensores no

puede ser fiable y totalmente optimizado.

• Por lo general, el modelado del entorno conlleva mucho tiempo y tiene un coste

muy alto, pero ninguno de los trabajos actuales propone algun metodo para

modelar el entorno de despliegue 3D con eficiencia y precision. Por lo tanto,

muchos investigadores estan limitados por este problema y sus algoritmos solo

se pueden evaluar en el mismo escenario, sin la posibilidad de probar la solidez

y viabilidad para las implementaciones en diferentes entornos.

En esta tesis, se propone una nueva metodologıa de planificacion ası

como una herramienta inteligente de planificacion de redes de sensores

inalambricas para ayudar a los disenadores a planificar WSNs fiables de

una manera eficiente.

En primer lugar, se propone un nuevo metodo para modelar de manera

eficiente y automatica los ambientes interiores y exteriores en 3D. Segun

nuestros conocimientos hasta la fecha, esta es la primera vez que las ventajas del

algoritmo de �image understanding�se aplican para reconstruir automaticamente

los escenarios exteriores e interiores en 3D para analizar la propagacion de la senal y

viii

la planificacion de la red. Los resultados experimentales indican que la metodologıa

propuesta es capaz de reconocer con precision los diferentes objetos presentes en las

imagenes satelitales de las regiones objetivo en el exterior y de la planta escaneada

en el interior. Su mecanismo ofrece a los usuarios la flexibilidad para reconstruir

los diferentes tipos de entornos sin ninguna interaccion humana. De este modo se

reduce considerablemente el esfuerzo humano, el coste y el tiempo invertido en la

reconstruccion de una base de datos geografica con informacion 3D, permitiendo ası

que los disenadores se concentren en los temas de planificacion.

En segundo lugar, se ha desarrollado un motor de trazado de rayos (en

ingles ray tracing) eficiente para modelar con precision la propagacion de

la senal de radio y la senal de los sensores en el mapa 3D construido. El

motor contribuye a la eficiencia y la precision de los resultados estimados. Mediante

el uso de los conceptos de procesamiento de imagenes, incluyendo el algoritmo del

arbol kd para la division del espacio y el algoritmo �polar sweep�modificado, los

rayos se trazan de manera eficiente sin la deteccion de todas las primitivas en la

escena. El modelo de propagacion de radio que se propone no solo considera los

materiales de los obstaculos, sino tambien su ubicacion a lo largo de la ruta de

senal. La senal de los sensores de los nodos, que es sensible a los obstaculos, se ve

beneficiada por la deteccion de objetos llevada a cabo por el algoritmo de trazado

de rayos. El rendimiento de este metodo de modelado es robusto y preciso en

comparacion con los metodos convencionales, y los resultados experimentales indican

que esta metodologıa es adecuada tanto para escenas urbanas al aire libre como para

ambientes interiores. Por otra parte, se puede aplicar a cualquier comunicacion GSM

o protocolo ZigBee mediante la variacion de la frecuencia del modelo de propagacion

de radio.

En tercer lugar, se propone un metodo de planificacion de WSNs

para hacer frente a los desafıos mencionados anteriormente y desplegar

redes de sensores fiables de manera eficiente. Se modelan mas parametros

(conectividad, cobertura, coste, tiempo de vida, la latencia de paquetes y tasa de

caıda de paquetes) en comparacion con otros trabajos. Especialmente el metodo

de trazado de rayos 3D se utiliza para modelar el enlace de radio y senal de los

sensores que son sensibles a la obstruccion de obstaculos; el enrutamiento de la red se

construye utilizando el protocolo AODV; la longevidad de la red, retardo de paquetes

ix

y tasa de abandono de paquetes se obtienen a traves de la simulacion de eventos

practicos en el simulador WSNet, y segun nuestros conocimientos hasta la fecha, es

la primera vez que simulador de red esta implicado en un algoritmo de planificacion.

Por otra parte, se ha desarrollado un algoritmo de optimizacion multi-objetivo para

satisfacer las caracterısticas de las redes inalambricas de sensores. La capacidad

de proporcionar multiples soluciones optimizadas de forma simultanea permite a los

usuarios tomar sus propias decisiones en consecuencia, obteniendo mejores resultados

en comparacion con otros algoritmos del estado del arte.

iMOST se desarrolla mediante la integracion de los algoritmos

presentados, para ayudar de forma eficiente a los disenadores en

la planificacion de WSNs fiables para diferentes configuraciones. El

nombre abreviado iMOST (Intelligent Multi-objective Optimization Sensor network

planning Tool) representa una herramienta inteligente de planificacion de redes de

sensores con optimizacion multi-objetivo. iMOST contribuye en: (1) Operacion

conveniente con una interfaz de facil uso, (2) Reconstruccion eficiente y automatica

de una base de datos con informacion 3D y diseno rapido de objetos 3D para

ambientes interiores y exteriores, (3) Proporciona varias soluciones de despliegue

optimizadas para los multi-objetivo en 3D y permite a los usuarios configurar

las propiedades de red, por lo que puede adaptarse a diversas aplicaciones de

WSN, (4) las soluciones de implementacion en el espacio 3D y el correspondiente

rendimiento evaluado se presentan visualmente a los usuarios, y (5) El �Node

Placement Module�de iMOST esta disponible en lınea, ası como el codigo fuente

de las otras dos heurısticas de planificacion. Por lo tanto los disenadores WSN se

beneficiaran de esta herramienta para la construccion eficiente de la base de datos

con informacion del entorno, la planificacion practica y eficiente de WSNs fiables

tanto para aplicaciones interiores y exteriores. Con los codigos fuente abiertos, son

capaces de comparar sus algoritmos desarrollados con los nuestros para contribuir a

este campo academico.

Por ultimo, se obtienen resultados reales solidos tanto para la

planificacion de WSN en interiores y exteriores. Los despliegues se han

realizado tanto para ambientes de interior y como para ambientes de exterior

utilizando las soluciones de planificacion propuestas. Los resultados medidos

coinciden en gran medida con los resultados estimados. El algoritmo de planificacion

x

propuesto se adapta convenientemente al deiseno de redes de sensores inalambricas,

y ofrece flexibilidad para planificar los despliegues 3D a pequena y gran escala tanto

en interiores como en exteriores.

La tesis se estructura en 7 capıtulos. En el Capıtulo 1, se presentan las

aplicaciones de WSN y motivaciones de este trabajo, se revisan los algoritmos

y herramientas de planificacion del estado del arte, se presentan los retos y se

describe brevemente la metodologıa propuesta. En el Capıtulo 2, se presenta

la metodologıa de reconstruccion de entornos 3D propuesta y su rendimiento es

evaluado tanto para espacios exteriores como para espacios interiores. El motor de

trazado de rayos desarrollado y el metodo de modelado de propagacion de radio

propuesto se describen en detalle en el Capıtulo 3, evaluandose en terminos de

eficiencia computacional y precision. En el Capıtulo 4 se presenta el modelado

de los parametros importantes de las WSNs y el algoritmo de planificacion de

optimizacion multi-objetivo propuesto, el rendimiento se compara con los otros

algoritmos de planificacion descritos en el estado del arte. La herramienta inteligente

de planificacion de redes de sensores inalambricas, iMOST, se describe en el Capıtulo

5. En el Capıtulo 6 se llevan a cabo despliegues reales de acuerdo a las soluciones

previstas para los escenarios interiores y exteriores, se miden los datos importantes

y se analizan los resultados. En el Capıtulo 7 se concluye la tesis y se discute acerca

de los trabajos futuros.

xi

Contents

1 Introduction 1

1.1 WSN applications and deployment experiences . . . . . . . . . . . . 2

1.1.1 Military applications . . . . . . . . . . . . . . . . . . . . . . . 3

1.1.2 Environmental applications . . . . . . . . . . . . . . . . . . . 4

1.1.3 Health applications . . . . . . . . . . . . . . . . . . . . . . . . 7

1.1.4 Other applications . . . . . . . . . . . . . . . . . . . . . . . . 9

1.2 Challenges when deploying WSNs . . . . . . . . . . . . . . . . . . . . 10

1.2.1 Necessity of simulation . . . . . . . . . . . . . . . . . . . . . . 12

1.3 Introduction to the planning algorithms and tools . . . . . . . . . . . 13

1.4 Proposed methodology and work flow: Main contributions . . . . . 22

1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2 3D environment reconstruction method 27

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.1.1 3D environment reconstruction from Lidar systems . . . . . . 28

2.1.2 3D environment reconstruction from images . . . . . . . . . . 29

2.2 3D outdoor environment reconstruction . . . . . . . . . . . . . . . . 32

2.2.1 Proposed algorithm for outdoor environment reconstruction . 35

2.2.2 Image database . . . . . . . . . . . . . . . . . . . . . . . . . . 36

2.2.3 Image understanding and segmentation algorithm . . . . . . 37

2.2.4 Performance enhancement . . . . . . . . . . . . . . . . . . . . 44

2.2.5 Shape matching and vectorization . . . . . . . . . . . . . . . 55

2.3 Indoor environment reconstruction . . . . . . . . . . . . . . . . . . . 64

2.3.1 Image calibration and classification . . . . . . . . . . . . . . . 65

2.3.2 Thinning and feature points extraction . . . . . . . . . . . . . 66

2.3.3 Smoothing and vectorizing . . . . . . . . . . . . . . . . . . . 66

Contents

2.3.4 Demonstration and analysis . . . . . . . . . . . . . . . . . . . 68

2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

3 Ray-tracing engine and radio propagation modelling 73

3.1 Space division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

3.2 Polar sweeping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

3.2.1 Direct path . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

3.2.2 Reflection path . . . . . . . . . . . . . . . . . . . . . . . . . . 81

3.2.3 Diffraction path . . . . . . . . . . . . . . . . . . . . . . . . . 82

3.3 Measurements and experimental results . . . . . . . . . . . . . . . . 83

3.3.1 Outdoor RF propagation verification . . . . . . . . . . . . . . 83

3.3.2 Indoor RF propagation verification . . . . . . . . . . . . . . . 88

3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

4 Planning the WSN 101

4.1 Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

4.1.1 Star network . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

4.1.2 Tree network . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

4.1.3 Mesh network . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

4.1.4 Cluster network . . . . . . . . . . . . . . . . . . . . . . . . . . 105

4.2 Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

4.3 Introduction and modelling of important metrics . . . . . . . . . . . 106

4.3.1 Preliminaries and assumptions . . . . . . . . . . . . . . . . . 106

4.3.2 The cost of WSN . . . . . . . . . . . . . . . . . . . . . . . . . 109

4.3.3 Coverage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

4.3.4 Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

4.3.5 Lifetime, Packet latency and Packet drop rate . . . . . . . . . 118

4.4 The proposed multi-objective optimization methodology . . . . . . 121

4.4.1 Initialization of individuals . . . . . . . . . . . . . . . . . . . 124

4.4.2 Crossover and mutation . . . . . . . . . . . . . . . . . . . . . 125

4.4.3 Evaluation based on desirability models and constraints . . . 126

4.5 Experimental results and analysis . . . . . . . . . . . . . . . . . . . 127

4.5.1 The impact of maximum number of generation . . . . . . . . 127

4.5.2 Performance comparison with other heuristics . . . . . . . . . 130

xiii

Contents

4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

5 iMOST: an Intelligent Multi-objective Optimization Sensor

network planning Tool 135

5.1 Menu bar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

5.1.1 Image Processing Module . . . . . . . . . . . . . . . . . . . . 136

5.1.2 Environment property configuration . . . . . . . . . . . . . . 136

5.2 Toolbar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

5.2.1 Node deployment . . . . . . . . . . . . . . . . . . . . . . . . . 138

5.2.2 Network Planning Module . . . . . . . . . . . . . . . . . . . . 138

5.2.3 Ray-tracing Propagation Module . . . . . . . . . . . . . . . . 140

5.2.4 3D navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

5.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

6 Real measurements and results analysis 143

6.1 Aggregation mechanism of measured data . . . . . . . . . . . . . . . 143

6.2 Application interface . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

6.3 Indoor measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

6.4 Outdoor measurements . . . . . . . . . . . . . . . . . . . . . . . . . . 155

6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

7 Conclusions and future works 163

7.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

7.2 Future works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

7.3 Publications based on this work . . . . . . . . . . . . . . . . . . . . . 166

7.4 Implementation of this work . . . . . . . . . . . . . . . . . . . . . . . 168

Bibliography 169

xiv

List of Figures

Fig. 1.1 A typical WSN architecture . . . . . . . . . . . . . . . . . . . . 2

Fig. 1.2 Environmental applications . . . . . . . . . . . . . . . . . . . . 5

Fig. 1.3 Health care applications . . . . . . . . . . . . . . . . . . . . . . 8

Fig. 1.4 Smartsantander . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

Fig. 1.5 Libelium smart world . . . . . . . . . . . . . . . . . . . . . . . . 11

Fig. 1.6 An example of tool interface . . . . . . . . . . . . . . . . . . . . 17

Fig. 1.7 Demonstration of the single-hop solution . . . . . . . . . . . . . 18

Fig. 1.8 Demonstration of the multi-hop solution . . . . . . . . . . . . . 18

Fig. 1.9 The proposed methodology . . . . . . . . . . . . . . . . . . . . . 24

Fig. 2.1 Airborne laser scanning . . . . . . . . . . . . . . . . . . . . . . . 28

Fig. 2.2 DSM with the original image . . . . . . . . . . . . . . . . . . . 28

Fig. 2.3 The principle of visual hull reconstruction . . . . . . . . . . . . 30

Fig. 2.4 Example of space carving reconstruction . . . . . . . . . . . . . 31

Fig. 2.5 Example of image-based rendering . . . . . . . . . . . . . . . . 31

Fig. 2.6 Workflow of the work by Saxena et al. . . . . . . . . . . . . . . 32

Fig. 2.7 Google street view tools . . . . . . . . . . . . . . . . . . . . . . 34

Fig. 2.8 3D outdoor reconstruction . . . . . . . . . . . . . . . . . . . . . 36

Fig. 2.9 MSRC database . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

Fig. 2.10 CEIeurope database . . . . . . . . . . . . . . . . . . . . . . . . 39

Fig. 2.11 17D filter bank . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

Fig. 2.12 The process of image textonization . . . . . . . . . . . . . . . . 40

Fig. 2.13 Calculating feature response . . . . . . . . . . . . . . . . . . . . 42

Fig. 2.14 Multi-object recognition procedure. . . . . . . . . . . . . . . . . 44

Fig. 2.15 Comparison between k-mean and graphcut . . . . . . . . . . . . 45

Fig. 2.16 Example of sub-clustering based on connectivity property. . . . 47

List of Figures

Fig. 2.17 The comparison between the proposed algorithm and that of

Shotton et al. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

Fig. 2.18 The principle of shadow detection . . . . . . . . . . . . . . . . . 50

Fig. 2.19 The result of shadow detection . . . . . . . . . . . . . . . . . . 51

Fig. 2.20 Road segmentation and orientation estimation . . . . . . . . . . 52

Fig. 2.21 Clustering of orientations . . . . . . . . . . . . . . . . . . . . . . 53

Fig. 2.22 Comparison of orientations . . . . . . . . . . . . . . . . . . . . . 54

Fig. 2.23 The result of road detection . . . . . . . . . . . . . . . . . . . . 55

Fig. 2.24 Frequently seen facade shapes of buildings . . . . . . . . . . . . 57

Fig. 2.25 Hierarchical shape matching . . . . . . . . . . . . . . . . . . . . 59

Fig. 2.26 Successful result of shape matching . . . . . . . . . . . . . . . . 61

Fig. 2.27 Shape registration . . . . . . . . . . . . . . . . . . . . . . . . . . 62

Fig. 2.28 KML shape description shown on Google Earth . . . . . . . . . 64

Fig. 2.29 Image calibration and segmentation . . . . . . . . . . . . . . . . 65

Fig. 2.30 Thinning step and feature point extraction . . . . . . . . . . . . 66

Fig. 2.31 Smoothing and regularization. . . . . . . . . . . . . . . . . . . . 67

Fig. 2.32 Vectorization result. . . . . . . . . . . . . . . . . . . . . . . . . 68

Fig. 2.33 A wall is described by four vertexes. . . . . . . . . . . . . . . . 68

Fig. 2.34 Reconstructed 3D indoor map in different views. . . . . . . . . 69

Fig. 2.35 A toy example by using the downloaded map downloaded . . . 70

Fig. 3.1 Space division by kd-tree . . . . . . . . . . . . . . . . . . . . . . 76

Fig. 3.2 Conventional Polar sweep. . . . . . . . . . . . . . . . . . . . . . 77

Fig. 3.3 Polar sweep. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

Fig. 3.4 Direct path. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

Fig. 3.5 Reflection path searching. . . . . . . . . . . . . . . . . . . . . . 81

Fig. 3.6 Diffraction path searching. . . . . . . . . . . . . . . . . . . . . . 82

Fig. 3.7 Three different routes measured by COST231 group. . . . . . . 84

Fig. 3.8 Classification and radio propagation over Munich scenario . . . 85

Fig. 3.9 Simulation result for the first route METRO200. . . . . . . . 86

Fig. 3.10 Simulation result for the second route METRO201. . . . . . . 86

Fig. 3.11 Simulation result for the third route METRO202. . . . . . . . 87

Fig. 3.12 Four-layer architecture and physical view of the Cookie node . . 88

xvi

List of Figures

Fig. 3.13 ETRX2 ZigBee communication module on communication layer 89

Fig. 3.14 Radio pattern of antenna of ETRX2 module . . . . . . . . . . . 89

Fig. 3.15 Three scenarios for radio measurements . . . . . . . . . . . . . . 90

Fig. 3.16 Ray tracing demonstrations from different TXs and RXs . . . . 91

Fig. 3.17 Simulation result: example 1 . . . . . . . . . . . . . . . . . . . . 92

Fig. 3.18 Simulation result: example 2 . . . . . . . . . . . . . . . . . . . . 92

Fig. 3.19 Results and comparisons of Scenario A . . . . . . . . . . . . . . 93

Fig. 3.20 Results and comparisons of Scenario B . . . . . . . . . . . . . . 94

Fig. 3.21 Results and comparisons of Scenario C . . . . . . . . . . . . . . 95

Fig. 3.22 Demonstration of the toy example . . . . . . . . . . . . . . . . . 96

Fig. 3.23 Average time consumption of polar sweeping . . . . . . . . . . . 97

Fig. 3.24 Average time consumption of ray tracing . . . . . . . . . . . . . 98

Fig. 3.25 Polar sweeping with and without kd-tree traversing . . . . . . . 99

Fig. 4.1 Different topologies of WSN . . . . . . . . . . . . . . . . . . . . 103

Fig. 4.2 The searching of covered point . . . . . . . . . . . . . . . . . . . 107

Fig. 4.3 Products by Crossbow . . . . . . . . . . . . . . . . . . . . . . . 110

Fig. 4.4 BTnode rev3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

Fig. 4.5 Waspmote,Wismote and SEED-EYE . . . . . . . . . . . . . . . 112

Fig. 4.6 Tyndall mote . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

Fig. 4.7 Deployment cost configuration in vertical view . . . . . . . . . . 116

Fig. 4.8 Modelling strategy by using WSNet simulator . . . . . . . . . . 121

Fig. 4.9 Proposed planning algorithm strategy . . . . . . . . . . . . . . . 122

Fig. 4.10 Crossover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

Fig. 4.11 Mutation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

Fig. 4.12 Configuration of scenario CEI-UPM . . . . . . . . . . . . . . . . 128

Fig. 4.13 Desirability values . . . . . . . . . . . . . . . . . . . . . . . . . . 129

Fig. 4.14 Time consumption . . . . . . . . . . . . . . . . . . . . . . . . . 129

Fig. 4.15 Scenario East Lansing . . . . . . . . . . . . . . . . . . . . . . . 131

Fig. 4.16 Scenario Madrid . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

Fig. 5.1 The mainframe of the planning tool . . . . . . . . . . . . . . . . 137

Fig. 5.2 User command on constructing new map . . . . . . . . . . . . . 137

Fig. 5.3 Environment property setting dialog . . . . . . . . . . . . . . . 138

xvii

List of Figures

Fig. 5.4 Node property configuration dialog . . . . . . . . . . . . . . . . 139

Fig. 5.5 Node configuration dialog . . . . . . . . . . . . . . . . . . . . . 140

Fig. 5.6 Generated topology demonstration . . . . . . . . . . . . . . . . 140

Fig. 5.7 3D navigation for outdoor scenario . . . . . . . . . . . . . . . . 141

Fig. 5.8 3D navigation for indoor scenario . . . . . . . . . . . . . . . . . 142

Fig. 6.1 BKR2400 antenna. . . . . . . . . . . . . . . . . . . . . . . . . . 144

Fig. 6.2 Application interface. . . . . . . . . . . . . . . . . . . . . . . . . 146

Fig. 6.3 Indoor modelling by using iMOST . . . . . . . . . . . . . . . . 148

Fig. 6.4 User requirement over the indoor test. . . . . . . . . . . . . . . 149

Fig. 6.5 Topology comparison 1 . . . . . . . . . . . . . . . . . . . . . . . 149

Fig. 6.6 Topology comparison 2 . . . . . . . . . . . . . . . . . . . . . . . 150

Fig. 6.7 RSS comparison between real measurement and simulation result 153

Fig. 6.8 Remaining energy of N1, N8 and N14 . . . . . . . . . . . . . . . 153

Fig. 6.9 Comparing the packet delivery status: Indoor . . . . . . . . . . 154

Fig. 6.10 The sensed data of N4. . . . . . . . . . . . . . . . . . . . . . . . 155

Fig. 6.11 Outdoor modelling by using iMOST . . . . . . . . . . . . . . . 156

Fig. 6.12 Topology comparison: Outdoor . . . . . . . . . . . . . . . . . . 159

Fig. 6.13 RSS comparison between real measurement and simulation result 160

Fig. 6.14 Remaining energy of N4, N5 and N2 . . . . . . . . . . . . . . . 160

Fig. 6.15 Comparing the packet delivery status: Outdoor . . . . . . . . . 161

xviii

List of Tables

Table 1.1 Comparison of planning algorithms . . . . . . . . . . . . . . . 20

Table 2.1 Confusion matrix of classification result . . . . . . . . . . . . . 47

Table 2.2 Shape matching result . . . . . . . . . . . . . . . . . . . . . . 60

Table 3.1 Comparison of radio estimation result with other methods. . . 87

Table 3.2 Results comparison: Scenario A . . . . . . . . . . . . . . . . . 93

Table 3.3 Results comparison: Scenario B . . . . . . . . . . . . . . . . . 95

Table 3.4 Results comparison: Scenario C . . . . . . . . . . . . . . . . . 96

Table 3.5 Attenuation parameters of major objects indoors. . . . . . . . 98

Table 4.1 Important symbols . . . . . . . . . . . . . . . . . . . . . . . . 106

Table 4.2 Features of various platform. . . . . . . . . . . . . . . . . . . . 114

Table 4.3 Features of algorithms for comparison . . . . . . . . . . . . . . 130

Table 4.4 Results comparison for Scenario CEI-UPM. . . . . . . . . . . . 133

Table 4.5 Results comparison for Scenario East Lansing. . . . . . . . . . 133

Table 4.6 Results comparison for Scenario Madrid. . . . . . . . . . . . . 134

Table 6.1 Routing table format. . . . . . . . . . . . . . . . . . . . . . . . 145

Table 6.2 Packet format. . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

Table 6.3 Evaluated performance of the two candidates. . . . . . . . . . 150

Table 6.4 Neighborhood table and RSS comparisons: Indoor . . . . . . . 152

Table 6.5 Evaluated performance of the selected candidate. . . . . . . . 157

Table 6.6 Neighborhood table and RSS comparisons: Outdoor . . . . . . 158

Chapter 1

Introduction

Recent years have witnessed an increased interest in the use of Wireless Sensor

Networks (WSNs) in various applications such as environmental monitoring, factory

automation, habitat tracking, security surveillance, intelligent transportation and

smart cities. This technology has brought a lot of benefits to the users from both

research and industrial areas. Fig. 1.1 depicts a typical sensor network architecture.

It is envisioned that tens to thousands of miniaturized sensor nodes, which operate

on small batteries, will be deployed to operate autonomously to construct WSNs

in different types of environments. Sensor networks may consist of various types

of sensors such as seismic, low sampling rate magnetic, thermal, visual, infrared,

acoustic and radar, which are able to monitor a wide range of ambient conditions

including [1]: temperature, humidity, light condition, the presence or absence of

objects, mechanical stress levels on attached objects, and the mobility characteristics

such as speed and direction. The sensed data are collected and sent to a base

station directly or via multiple hops depending on the network topology and routing

protocols. In addition to the ability to probe its surroundings, each sensor node

has one or more onboard radios to communicate with other nodes through wireless

communication protocols such as ZigBeeTM [2], BluetoothTM [3] and Ultra-wideband

(UWB)[4] among others. Therefore the combinations of micro-sensing and wireless

communication offers a huge number of possibilities of WSN applications.

The categorization of the applications of WSNs differs by different researches.

Some works categorize the applications into military, environment, health, home

and other commercial areas [5]; whereas some classify them into two categories

[6]: monitoring and tracking. It is possible to expand this classification with other

categories such as event detection and spatial process estimation [7].

In this chapter, we enumerate the WSN applications based on the monitoring

and tracking frame with the items of different implemented areas as mentioned in [5],

Chapter 1. Introduction

Base station

Applications

Figure 1.1: A typical WSN architecture.

the deployment experiences of some applications are reviewed as well. Afterwards,

the challenges of deploying a WSN are identified to state out our motivations and

targets. The main frame of the proposed planning methodology and planning tool

are briefly described at the end of this chapter.

1.1 WSN applications and deployment experiences

The origin of WSNs can be traced back to the 1950s during the cold war between

Soviet Union and the United States. The SOund SUrveillance System (SOSUS) was

developed by the United States Military to detect and track Soviet submarines. In

order to echo the investments made in the 1960s and 1970s to develop the hardware

for today’s Internet, the United States Defense Advanced Research Projects Agency

(DARPA) launched the Distributed Sensor Network (DSN) program in 1980 to

formally explore the challenges in implementing distributed wireless sensor networks.

With the birth of DSN and its progression into academia, the governments

and universities eventually show their interests in this topic and promote the

research atmosphere to improve the performance of WSN and explore new areas

of applications using WSNs, such as air quality monitoring, forest fire detection,

natural disaster prevention and structural monitoring. Then there arises a strong

2

1.1. WSN applications and deployment experiences

demand in the market as researchers made their steps into the corporation with

technology giants of the day, such as IBM, CISCO and Bell Labs, they began to

explore and expand the use of WSNs in our daily life, such as security surveillance,

health monitoring and smart life. Lessons and extensive experiences are learnt

during designing and constructing the wireless sensor networks, which continuously

provide challenging issues and attracts more interests in this area.

1.1.1 Military applications

The C4ISRT (Command, Control, Communications, Computers, Intelligence,

Surveillance and Reconnaissance) systems integrate WSNs to realize military

monitoring and tracking. The systems can be used to monitor friendly forces,

equipment and ammunition. Every troop, vehicle, equipment and critical

ammunition can be attached with small sensors to report the statuses. Those reports

are gathered in sink nodes or be forwarded to the upper levels of the hierarchy and

aggregated with the data from other units at each level, and sent to the troop

leaders at the end. They are also implemented for battlefield surveillance where

critical terrains, approach routes, paths and straits can be rapidly covered with

sensor networks, to closely monitor the activities of the opposing forces. As a result,

the deployment efficiency and reliability are very demanding for such applications to

allow WSNs being constructed expeditiously and able to cover the area of interest.

There are other military applications developed for targeting and tracking:

Sensor networks can be incorporated into guidance systems of the intelligent

ammunition, they can also be attached to soldiers to track their mobility and

locations. In [8], the invasion of individual enemy soldiers are detected by using

unattended acoustic and seismic sensors in the protecting military sites or buildings.

The Early attack reaction sensor (EARS) [9] is a man-wearable gunshot system. It

uses passive acoustic sensing system with small microphone array to detect gunshots

(muzzle blast and/or shockwave) and provide relative azimuth and range information

of the shot origin to the user. It has been tested in both open field and military

operations in urban terrain (MOUT) environment and has provided useable bearing

and range information against the firing positions. Some systems also perform

localization such as [10, 11] to protect soldiers from potential menace by blasts

3


and snipers. Such applications should pay sufficient attentions to the connections

of nodes.

In [12], hazardous chemicals are detected and identified by their unique infrared

absorption signatures. The application is for deployment on expendable unmanned

aerial vehicles in a nadir-viewing configuration from an altitude of 300 m while

traveling at an air speed of 96 km/h. The applications of monitoring missile

environment target at dramatically increasing missile active service life, saving

millions of dollars and reducing the number of missiles needed. The challenge of such

real-time monitoring systems is to collect and store data on environmental shock

with high speed and observe vibration (up to 100 g) in missile canisters without

electrical hazards. An optical sensor system capable to monitor shock and vibration

in missile canisters in three dimensions at high speed (5 kHz) is proposed in [13].

1.1.2 Environmental applications

Environmental monitoring has been studied for a long time. It includes the detection

and reaction towards natural disaster such as earthquake and avalanche, the

detections on climate change and pollution, the tracking and observations of animal

behaviors. The old mechanisms recorded data at specific intervals and required

human intervention to download them. The employment of WSN technology in

this field shortens the time and reduces efforts on data aggregation issue, and most

importantly it is capable to monitor/track a huge quantity of the environmental data

without disturbing the natural environment too much compared with the traditional

approach.

W-TREMORS [14] is a platform with high-frequency distributed data acquisition

ability, and it is designed for earthquake engineering and structural monitoring. By

adopting a novel communication protocol together with the developed software,

it is able to test the shaking with inexpensive hardware resources. Wong et al.

from UC Berkeley built a wireless sensor seismic response monitoring system based

on MICA2 motes [15]. The system was tested by using a reinforced concrete

bridge column (Fig. 1.2(a)). However, the accelerometer and the analog-to-digital

converter implemented in the MICA2 motes do not possess the fidelity required

for structural state evaluation. Both works mentioned about resolution problems,

due to quantization significantly affects the low level acceleration readings. Besides,

4


(c)

(a)

(b)

Figure 1.2: Environmental applications. (a) The test of WSN on specimen [15], (b)ZebraNet [16], (c) The volcanic monitoring system[17].

due to the interferences in radio transmission with multiple motes attempting to

communicate simultaneously, the communication packet loss is quite pronounced at

high sampling rates.

ZebraNet system [16] shown in Fig. 1.2(b) is a mobile wireless sensor network

used to track animal migrations. ZebraNet is composed of sensor nodes built into the

zebra’s collar. The node consists of a 16-bit TI microcontroller, 4 Mbits off-chip flash

memory, a 900 MHz radio, and a GPS unit. Positional readings are taken using the

GPS and sent multi-hop across zebras to the base station. The goal is to accurately

log each zebra’s position and use them for analysis. In the demonstration, a total of

6-10 zebra collars were deployed at the Sweetwaters game reserve in central Kenya to

study the effects and reliability of the collars and to collect movement data. After

deployment, the biologists observed that the collared zebras were affected by the

collars. They observed additional head shakes from those zebra in the first week.

After the first week, the collared zebra showed no difference than the uncollared

zebra. A set of movement data were also collected during this study, from which

the biologists can better understand the zebra movements.

5


In volcanic monitoring, the challenges of a WSN application for data collection

include reliable event detection, efficient data delivery, high data rates, and sparse

deployment of nodes. Given these challenges, a network consisting of 16 sensor nodes

was deployed on Volc�n Reventador in northern Ecuador [17]. Each sensor node is

a T-mote sky [18] device equipped with an external omni-directional antenna, a

seismometer, a microphone, and a custom hardware interface board. Overall, the

system performed well in this study. In the 19 days of deployment (Fig. 1.2(c)),

the network observed 230 eruptions and other volcanic events. About 61% of the

data was retrieved from the network due to short outages in the network from

software component failure and power outage. Therefore, the node devices in this

type of application should be smaller, lighter, and consume less power to facilitate

distribution and prolong lifetime of the network.

Macroscope of redwood [19] is an experimental application of a WSN that

monitors and records the redwood trees in Sonoma, California. Sensor nodes are

placed at different heights of the tree. Air temperature, relative humidity, and

photo-synthetically-active solar radiation are measured by each sensor node. Plant

biologists track changes of spatial gradients in the microclimate around a redwood

tree and validate their biological theories.

Sensor nodes may be deployed in a forest strategically to relay the exact origin

of the fire to the end users before the fire is spread uncontrollable. Forest fire

monitoring systems require a large amount of sensor nodes being deployed and

integrated using radio frequency/optical systems. Also, they may be equipped with

effective energy harvesting methods [20], such as solar cells, because the sensors

may be left unattended for months and even years. The sensor nodes collaborate

with each other to perform distributed sensing and overcome obstacles, such as

trees and rocks, that block line of sight of sensors. DIMAP-FactorLink provides

another example of forest fire detection. It has developed and integrated a forest

fires detection system using the products of Libelium [21]. The covered area of the

system was about 210 hectares in the north Spain region. Therefore in this type

of application, we should focus not only on the coverage and communication issues,

but also on the lifetime and cost.

Biocomplexity mapping of the environment is done at the James Reserve in

Southern California [22]. Three monitoring grids with each having 25-100 sensor

6


nodes are implemented for fixed view multimedia and environmental sensor data

loggers. The ALERT system [23] is another example deployed in the US for flood

detection. The types of sensors deployed in the system include rainfall, water level

and weather sensors.

The authors in [24] developed an underwater WSN application platform for

long-term monitoring of coral reefs and fisheries. The deployed sensor network

consists of static and mobile underwater sensor nodes. The nodes communicate with

each other via point-to-point links using high speed optical communications with an

acoustic protocol integrated in the TinyOS protocol stack. They have several types

of sensors, including temperature and pressure sensing devices and cameras. Mobile

nodes are needed to locate and move above the static nodes to collect data and

perform network maintenance functions for deployment, re-location, and recovery.

WSNs applications on precision agriculture have recently appeared. For

instances, the experiences in the design, development and deployment of a WSN

are described in [25] to improve water use efficiency for pasture production. Sensor

pods should be designed carefully in this application to withstand seasonal weather

changes and are resistant to damages that may be inflicted by cattle in the field.

Temperature and humidity were measured using the Tmote Sky’s on-board sensors.

70 sensor pods were deployed in December of 2007, at the TIAR (Tasmanian

Institute for Agricultural Research) Elliott Research Farm and managed to gather

correct data from the field. The iCubes were designed with low-cost humidity sensors

to monitor the soil wetness in the work of [26] and the authors in [27] developed

a monitoring system by using TelosB wireless sensor nodes to acquire data such as

temperature, humidity, illumination and voltage. In this application, a web-based

platform integrated with Google Maps was developed to release the greenhouse

environmental status and provide real-time voice and SMS alarm service.

1.1.3 Health applications

Health applications can be categorized into activities of daily living monitoring,

fall and movement detection, location tracking, medication intake monitoring, and

medical status monitoring [28]. The Ultra Badge System [29] is one example of

location tracking application that is used in a hospital setting. In Ultra Badge, a

3D tag system is designed to localize the patients. When a patient is in a specific area

7


where a fall is most likely to occur, the system alerts the caregivers beforehand. The

Ultra Badge System consists of ultrasonic receivers embedded in the environment

and wireless ultrasonic emitters placed on objects as depicted in Fig. 1.3(a). The

positions of the receivers are fixed and known beforehand, therefore the emitters

can be positioned by using the multi-lateration technique. Two subsystems were

developed for real implementations: the wheelchair locator and the ultrasonic radar.

In the former subsystem, the nurse is notified when a patient uses a wheelchair

approaching a ’detection area’, where a fall is most likely to occur. The latter

subsystem aims to monitor the activities of the patients in their beds by using

ultrasonic pulses.

(a) (b)

Figure 1.3: Health care applications. (a) The Ultra Badge System[29], (b) TheiPackage of [30].

As patients may get allergic by taking some medications or risk of life with wrong

dosage, if sensor nodes can be used in medication monitoring, the chance of getting

and prescribing the wrong medication to patients can be minimized. An intelligent

packaging prototype (iPackage) is developed by Pang et al.[30]. The system is

capable of both remote medication intake monitoring and vital signs monitoring.

It uses an array of Controlled Delamination Material (CDM) films along with the

control circuits. The CDM film is a three-layer foil composed of aluminum bottom

and top layers and an adhesive middle layer made of electrochemical epoxy. When

a voltage higher than a particular threshold is applied on the bottom layer and top

layer, an electrochemical reaction occurs in the middle layer. When the voltage is

8


applied for a certain amount of time, the epoxy layer is destroyed and delaminated.

Therefore, the iPackage sealed with a CDM film can only be opened by the special

control appliance which also enables the control of the dosage. The identification of

the correct pill is accomplished by RFID. The prototype design of CDM and tagged

capsule package is depicted in Fig. 1.3(b).

1.1.4 Other applications

In the recent years, the concept of smart cities are proposed as the next stage in the

process of urbanisation. As a city is a system of systems, the more we understand

how those systems interact and share information, the better people can be helped

to make decisions and to make the city better. The smart city can be identified

along eight main axes or dimensions: smart environment, smart energy and water,

smart transportation, smart education, smart healthcare, smart public safety, smart

buildings and urban planning and smart government.

Beyond the previously mentioned applications for smart environment, smart

energy and water and smart healthcare, there exist smart transportation applications

such as the PGS smart parking system [31], which is developed in ICU

Korea based on a new T-Sensor hardware. And the California Partners for

Advanced Transportation TecHnology (PATH) was established, with the mission

to develop solutions to the problems of California’s surface transportation systems

through cutting edge research, PATH research is divided into three program

areas: Transportation Safety Research, Traffic Operations Research and Modal

Applications Research. The traffic surveillance by WSNs was developed based on

MICA2 DOTS. It was used to control traffic signal, on-ramp metering system to

regulate the flow of traffic on freeway entrance ramps using traffic signals, the system

was able to reduce delay by 102 million person-hours in 2003. The parking guidance

and information system (PGIS), road condition sensing modality were also developed

in PATH research.

SmartSantander project [32] proposes a city-scale experimental research facility

to support typical applications and services for a smart city. This unique

experimental facility is sufficiently large, open and flexible to enable horizontal

and vertical federation with other experimental facilities. The project envisions a

9


Figure 1.4: Outdoor parking and environmental monitoring deployed architectureof Smartsantander project.

deployment of 20,000 sensors in Belgrade, Guildford, L�beck and Santander (12,000

nodes, Fig. 1.4), which exploits a large variety of technologies.

One of the famous sensor nodes producers, Libelium, creates the inspirational

and market research documents on 50 sensor applications. It has comprised and

concluded necessary applications in a infographic (Fig.1.5) which combines Smart

Cities, Internet of Things (IoT) and other sensing applications to construct a smarter

world.

1.2 Challenges when deploying WSNs

One of the major challenges in designing WSNs is the support of capturing and

gathering data requirements while coping with the computation, energy, sensing

ability and communication constraints. As a result, careful node placement can be

a very effective optimization mean for achieving the desired design goals.

In order to prolong the lifetime of WSN, energy conservation methods must be

taken, scheduling and data aggregation are among the commonly used methods.

Scheduling conserves energy by turning off the sensor whenever possible. While

data aggregation tries to conserve energy by reducing the energy used in data

transmission, efficient routing and topology construction can significantly reduce the

energy consumption due to data aggregation. As the energy harvesting researches

blossom nowadays, with the aim to tackle the energy constraint issues for different

10

1.2. Challenges when deploying WSNs

Figure 1.5: Libelium smart world.

areas, it will be able to solve the limited battery resource issue and allow the WSN

applying to more areas.

Connectivity and coverage problems are caused by the limited communication

and sensing ability of sensor nodes. To solve both problems, the solution lays in how

the sensors nodes are positioned with respect to each others. Coverage problem is

regarding how to guarantee that each of the points to be monitored is covered by

the sensors in the region. It is a trade-off problem, in maximizing coverage with low

cost: the sensors need to be placed not too close to each other so that the sensing

capability of the network is fully utilized while the cost is minimized; at the same

time, they must not be located too far away from each other to avoid the formation

of coverage holes (area outside sensing range of sensors). On the other hand, it is

also a trade-off problem in terms of connectivity and cost: the sensor nodes need to

be placed close enough so that they are within each other’s communication range

and ensure the connectivity of WSN with robustness and reliability.

11


1.2.1 Necessity of simulation

The emergence of wireless sensor networks brought many new challenging issues

to WSN designers. Traditionally, the three main techniques for analyzing the

performance of wired and wireless networks are analytical methods, computer

simulation, and physical measurement. Due to many constraints imposed on

sensor networks, as mentioned above, the energy limitation, sensing ability

and communication constraints, the deployment of sensor networks tends to be

quite complex and requires mastering of cross knowledge on networking, radio

propagation and embedded system. Furthermore, although the aforementioned

WSN applications have been implemented, there is still a lack of analysis and

experiences to guide WSN designers to construct WSN topology successfully without

many trials. Therefore, it appears that simulation is a feasible approach to the

quantitative analysis of the performance of wireless sensor networks.

The authors of [33] introduce their deployment experiences of various WSN

applications and recommend that simulation should come first before real

deployment. But for getting realistic results, one must have a realistic simulation

environment, in which all parameters concur to an accurate description of the

environment, platform and operation. In order to achieve realistic results, one can

barely rely on the pre-defined parameters found in the literature.

Currently, there are many conventional open source and freeware WSN

simulation tools that are publicly available or in academic research use. Some

examples are NS-2 [34], OMNeT++ [35], Worldsens [36] and TOSSIM [37], in which,

the network protocols can be programmed and configured according to users’ desires

and provide the convenience to simulate and evaluate performances at protocol

levels. However, radio signal propagation is a very complex phenomenon since

it is three-dimensional and influenced by many disturbances that are caused by

the environment. Multi-path fading and attenuation directly contribute to the

reliability and range of the wireless network. The current network simulators

all use very simple physical channel and environment modeling. As default, the

propagation models in these simulators are based on predefined empirical functions

or assume as line-of-sight connection where no obstacles between the transmitter and

receiver. Despite their powerful ability in validating WSN performance at protocol

functionality level, they are not able to consider the impacts of realistic environments

12

1.3. Introduction to the planning algorithms and tools

on topologies, protocols and deployments, and as a result, the performance of

designed WSN can not be practically estimated, which might significantly influence

the real implementation.

Besides the network simulators for validating protocol levels, several other tools

have been developed for radio propagation level or pre-deployment level to tackle

different challenging issues for WSNs. In [38], the authors introduced several

state-of-the-art radio propagation simulators, such as the EDX Signal pro [39],

Winprop (AWE) [40] and CINDOOR [41], all of which are featured with 3D indoor

modeling and ray-tracing propagation modeling in the simulation. The simulation

and modeling of radio propagation for WSN can significantly reduce the work effort

and costs to ensure the connectivity of WSNs. For obtaining more accurate results,

the suitable radio simulator for WSNs is envisaged to support the importation of 3D

indoor and outdoor environment model, the radio and antenna should be definable,

the radio algorithm should be able to detect the multi-path effects including the

direct path, reflections and diffractions.

After investigating on the deployment issues and performance evaluation issues

of WSNs, two questions arise:

• Why can’t a good radio frequency simulator be used in collaboration with

a network simulator to provide practical WSN performance analysis for any

specific 3D environment and application?

• Once provided with the evaluated performance of WSN topology, is there any

smart tool to assist WSN designers improving the overall performance of WSN

automatically and efficiently instead of manually adjust the designed topology

and estimate repeatedly through network simulator?

The answers to those questions are directedly pointed to smart planning tool for

WSNs, which is desired and strongly demanded by the WSN designers.

1.3 Introduction to the planning algorithms and tools

There are several WSN planning algorithms and tools developed in recent years. The

3D indoor planning heuristic (LowCost) [42], to the best of our knowledge, is the first

indoor 3D WSN deployment heuristic that considers impacts of obstacles on sensing

13


signal and radio communication. It consists of two steps: Provided a 3D indoor

environment model with furniture and obstacles recorded, the first step calculates

the coverage to deployment cost ratio for all the candidate points in the deployable

area. Sensor nodes are iteratively put at the point with the maximum coverage to

deployment cost ratio, so that the target region is covered with the minimum sensor

node cost after this step. Then the connectivity of the deployed nodes is checked

in the second step. The authors consider two options to satisfy the connectivity of

WSN, the prior one is realized by moving the unconnected node towards the closest

connected node without influencing the sensing coverage of the first step; otherwise,

if the preferential option is not applicable, extra sensor nodes will be added along

the line between the unconnected node and the closest connected node. Note that

despite this approach manages to cover the sensing area with the ”minimum cost”,

the connectivity of the WSN is ensured by simply moving or placing extra nodes

without carefully selecting optimal positions to decrease the hardware cost, improve

the link quality or prolong the network lifetime. Moreover, although the modelling

of the sensing signal considers obstacles, the radio propagation model is too simple

because the communication links are established only between line-of-sight (LOS)

nodes, which is obviously not true in the real-world propagation.

The MOGA algorithm [43] employs multi-objective genetic algorithm, which is

proved to be efficient in solving NP-hard problem, to evolve the decision. Topology

solution for the same network varies at different runs which provides more options

than the deterministic approach of LowCost. However it focuses on maximizing

the sensing coverage and prolonging the network lifetime with a predetermined

number of nodes, as a result the hardware cost can not be optimized. Moreover

the modeling of radio signal and sensing signal are based on ideal disc model thus

it is not environmental sensitive.

The previous two methods are developed for planning the homogeneous WSNs.

However, in many prototypical systems available today, sensor networks normally

consist of a variety of different devices. Nodes may differ in the type and number of

attached sensors [32, 44]; some computationally more powerful �compute�nodes

may collect, process, and route sensory data from many more limited sensing nodes

[45].

14


There are several works focused on planning heterogeneous network, such as in

[46], the relay node placement problem for WSNs is concerned as placing a minimum

number of relay nodes into a WSN to meet certain connectivity or survivability

requirements. In that work, the authors assume that there may be some physical

constraints on the placement of relay nodes and they study constrained versions

of the relay node placement problem, where relay nodes can only be placed at a

set of candidate locations. In the connected relay node placement problem, they

want to place a minimum number of relay nodes to ensure that each sensor node is

connected with a base station through a bidirectional path. In the survivable relay

node placement problem, they want to place a minimum number of relay nodes to

ensure that each sensor node is connected with two base stations (or the only base

station in case there is only one base station) through two node-disjoint bidirectional

paths. For each of the two problems, they discuss its computational complexity and

present a framework of polynomial time O(1)-approximation algorithms with small

approximation ratios. Numerical results show that their approximation algorithms

can produce solutions very close to optimal solutions. The authors of [47] propose

an approximation algorithm to find a feasible solution for relay node placement

to deploy a minimum set of relay nodes in such a fashion that each sensor node

must have at least one relay node within its one hop distance and all deployed relay

nodes eventually form a connected network among themselves including one or more

base-stations. The work reveals an approximation algorithm that runs in O(n2) time

complexity, to find a feasible solution for above challenge. The authors also describe

a framework to solve the above problem in non-convex shaped deployment region.

[48] targets at providing hight efficiency and QoS and it presents a polynomial-time

algorithms which is QoS-aware relay node placement using minimum Steiner tree

on Convex hull.

The work in [49] proposes Multiple-Objective Metric (MOM) for base station

placement in wireless sensor networks to fairly increase various properties. It

considers four different metrics for base station placement in WSNs. First, the ratio

of sensor nodes which can communicate with a base station via either single-hop

or multi-hop represents the coverage of sensor nodes. Second, the average ratio of

connected sensor nodes after the failure of base stations represents the fault tolerance

of a network. Third, the average distance between sensor nodes and their nearest

15


base station represents the energy consumption of a network. However, as discussed

before, not only the distance but also the obstacles lead to attenuation of received

signal strength. Moreover, more energy is consumed at nodes with larger degree, as

a result the energy consumption is not practically modelled by this work. Fourth,

the standard deviation of the degree of base stations represents the average delay of

a network due to congestion. The limitation of this algorithm is that sensor nodes

should be pre-located by designers, which neither guarantees the sensing coverage

without expert experience nor allows optimizing the hardware cost for WSN.

The authors in [50] develop a tool that integrates a developed 3D indoor

deployment heuristic together with NS-2 simulator to assist designers deploying and

analyzing the performance of network. They propose a heuristic that minimizes

hardware cost while satisfying requirements on coverage and connectivity. The

network topology is constraint to the type of cluster tree and three different devices

are provided: the coordinator, router and sensor. Sensors can only communicate

with routers and coordinator. The heuristic considers radiation pattern of antenna

as well as the effects of obstacles by using accurate ray-tracing algorithm. Once

the topology is generated, the integrated NS-2 simulator is driven to simulate the

packet drop rate and latency, and the results are demonstrated to users. The

merit of this method is the integration of authorized network simulator to evaluate

the performance of generated topology which provides a much more practical

implication on packet delivery performance to designers. However, as the evaluation

from NS-2 has no contribution to improve the generated topology, the proposed

deployment heuristic should be run several times so that by a certain chance,

designers can observe a satisfied solution with low cost, low drop rate and latency.

The user interface of this work is shown in Fig. 1.6. It allows users prosecute

many configurations including map, node properties, topology constraints and

environment types. The generated topology can be shown and results evaluated

by NS-2 are reported on the interface.

Another state-of-the-art method for deploying relay node and sink node for

indoor environment is proposed in [51, 52], the tool allows users defining the

node demand zones, power source, sensing interval and transmission delay. By

encapsulating those metrics into a complete requirements model, the tool optimizes

the infrastructure of WSN and maximizes the utility function which provides

16


Figure 1.6: The interface and demonstration of the work in [50].

a normalized equation that observes the coverage, link quality, lifetime and

infrastructure cost. Fig. 1.7 and Fig. 1.8 are two examples of generated solutions

for single-hop and multi-hop topologies respectively.

The lifetime (L) of sensor node is considered in that work and is modelled by

(1.2), The electric charge of a sensor node EC, expressed in mAh, is calculated

according to (1.1) where Ia and Is are the power consumption in active state and

sleep state respectively. ta and ts represents their time durations in a node interval.

The current capacity of the battery CC is expressed in mAh.

EC =3600

ta + ts× (ta × Ia + ts × Is) (1.1)

L =CC

EC(1.2)

As it can be noticed from the formulations, the model of lifetime only considers the

impacts of different states and their corresponding time durations, while the number

of packets forwarded for other nodes are ignored which is not realistic especially for

multi-hop topology. Besides, the authors did not consider the impacts of packet

delivery ability (latency and drop rate) to ensure a more reliable WSN topology.

17


Sensor nodes in this work should also be pre-determined by users and therefore

node locations and cost are not optimized.

Figure 1.7: Demonstration of the single-hop solution by the work in [51].

Figure 1.8: Demonstration of the multi-hop solution by the work in [51].

18


Some works are developed to tackle the modelling of sensing signal and radio

signal to make the deployment algorithm more practical and accurate. The authors

in [53] develop a probabilistic sensing model for sensors with line-of-sight-based

coverage (e.g. cameras). The probabilistic sensing model takes into consideration

sensing capacity probability as well as critical environmental factors such as terrain

topography. Besides, they also implement several optimization schemes for sensor

placement optimization. Sensor deployment in network-structured environments

is studied in [54] and it aims to achieve k-coverage while minimizing the number

of sensor nodes. The coverage problem of wireless sensor networks for the rolling

terrains is studied in [55] to derive the general expression of the expected coverage

ratio for regular terrains and irregular terrains.

To enhance the WSN lifetime, the authors in [56] propose a deployment

strategy with a non-uniform deployment method and an alternative duty mode

to balance the energy consumption of sensor nodes in chain-type WSNs. To make

the deployed network resilient to faults caused by communication errors, unstable

network connectivity, and sensor faults, the authors in [57] present an approach,

called FTSHM (fault tolerance in Structural health monitoring (SHM)), to repair

the network with redundant backup nodes and guarantee a specified degree of

fault tolerance. FTSHM searches the repairing points in clusters and places a

set of backup sensors at those points by satisfying civil engineering requirements.

FTSHM also includes a SHM algorithm suitable for decentralized computing in

energy constrained WSNs, with the objective to guarantee that the WSN for SHM

remains connected in the event of a sensor fault thus prolonging the WSN lifetime

under connectivity and data delivery constraints. Table. 1.1 provides a comparative

summary of the characteristics of the static node placement mechanisms discussed

in this section.

Optimal node placement is a very challenging problem that has been proven

to be NP-Hard for most of the formulations of sensor deployment [59, 60, 61].

The deployment of WSNs should satisfy the requirements on the 3D sensing

coverage, guaranteeing connectivity, prolonging the network longevity and reducing

the hardware cost.

However the existing planning algorithms and tools are to some extent have

serious limitations to practically design reliable WSN topology. Only a few of

19

Chapter 1. IntroductionTable

1.1:

Compariso

nof

planningalgorith

ms

Paper

Applic

atio

nSpace

Deployment

Node

type

Objectiv

es

Constra

ints

[42]

Indoor

Gen

eric3D

Determ

inistic

Sensor

Deploy

mentcost

Hardware

cost

-

[43]

Generic

2D

Random

Sensor

Covera

ge,

Life

time

Fixed

nodecount

[46]

Gen

eric2D

Determ

inistic

Relay

node

Connectiv

ityLifetim

eHardware

cost

Place

ofrelay

node

[47]

Generic

2D

Random

Relay

node

Connectiv

ity,Hardware

cost

Onesen

soratta

ched

toonerelay

node

[49]

Generic

2D

Random

Base

statio

nConnectiv

ity,Fau

ltToleren

ceEngerg

yconsumptio

nCongestio

n–

[58]

Generic

1D

Determ

inistic

Sensor

Lifetim

e,Cost

Linearalig

ndeploy

ment

[48]

Gen

eric2D

Determ

inistic

Relay

node

QoS,Cost

Con

vexhull

[53]

Generic

3D

Determ

inistic

Sensor

Covera

ge

Cost

–

[54]

Gen

ericNetw

ork

Stru

cture

Determ

inistic

Sen

sor,S

ink

Covera

ge

Data

fidelity

–

[55]

Outdoor

Gen

eric3D

Random

Sensor

Covera

ge

Cost

Sphere

sensin

gmodel

[56]

Generic

1D

Determ

inistic

Sensor

Lifetim

e–

[57]

Stru

cture

Monito

ring

3D

Determ

inistic

Sensor

Fau

lttolera

nce

–

[50]

Indoor

3D

Determ

inistic

Sensor

Coordinator

Router

Cost

Covera

te,Connectiv

ity

[51,

52]

Indoor

3D

Determ

inistic

Relay

node,

Sink

Linkquality,L

ifetime

Infra

structu

recost

Covera

ge,

Connectiv

ity

20


them tackle the 3D deployment issue, and an overwhelming number of works are

proposed to place devices in 2D scheme. Without considering the full dimension, the

impacts of environment to the performance of WSN are not completely studied, thus

the values of evaluated metrics such as connectivity and sensing coverage are not

sufficiently accurate to make proper decision. Even fewer planning methods model

the sensing coverage and radio propagation by considering the realistic scenario

where obstacles exist. Radio signals propagate with multi-path phenomenon

in the real world, in which direct paths, reflected paths and diffracted paths

contribute to the received signal strength. Besides, obstacles between the path

of sensor and objects might block the sensing signals, thus create coverage hole

in the application. None of the planning algorithms model the network longevity

properly and practically. They often employ unilateral and unrealistic formulations.

The optimization targets are often one-sided in the current works. Without

comprehensive evaluation on the important metrics, the performance of planned

WSNs can not be reliable and entirely optimized.

The diversity and variety of the deployed environment of WSNs have significant

impacts on radio communication and sensing coverage. As a result, it is crucial

for WSN designers having accurate environment model to prosecute realistic radio

propagation and coverage estimation. However, in conventional approaches of

constructing 3D environment model, people have to use a third party CAD tool

or geometric scripts to reconstruct the target environment database. The prior

knowledge of the detailed geographical information can only be obtained via manual

measurements or from the city hall. Instead of doing so, people can buy such

services from professional 3D modelling agencies. When the application scenario

gets larger, the cost grows higher. Therefore, either time and efforts or money

is needed in conventional approaches to construct 3D environment models. While

none of the current network planning works or radio propagation researches figure

out any method to model the 3D deployment environment efficiently and accurately.

Many researchers are trapped by this issue and their algorithms/models can only be

evaluated always in the same scenarios, without the possibility to test the robustness

and feasibility for implementations in different environments.

21


1.4 Proposed methodology and work flow: Main

contributions

Motivated by all the reasons and challenging issues mentioned in the previous

sections, this work dedicates to explore and develop state-of-the-art planning

methodology and smart planning tool to assist WSN designers efficiently designing

reliable WSN deployments.

• A new method is proposed to efficiently and automatically model the 3D indoor

and outdoor environments. The advantages of image understanding algorithm

are applied to automatically recognize objects from the satellite images of

the outdoor target regions and from the scanned floor plan of indoor area,

thereafter 3D outdoor/indoor scenarios are reconstructed automatically and

efficiently for signal propagation and network planning purpose. Its mechanism

offers users a flexibility to reconstruct different types of environment without

any human interaction. Thereby it significantly reduces human efforts, cost

and time and allows WSN designers concentrating on the planning issues.

• An efficient ray-tracing engine is developed to accurately and practically model

the radio propagation and sensing signal on the constructed 3D map. By using

the kd-tree space division algorithm and modified polar sweep algorithm, the

signal rays are traced efficiently without detecting all the primitives in the

scene. The proposed radio propagation model, which emphasizes not only

the materials of obstacles but also their locations along the signal path, is

applied to compute the received signal strengths for all candidate receivers.

The sensing signal of sensor nodes is tracked by taking advantage of this

obstacle sensitive approach.

• WSN planning method is proposed to tackle the above mentioned challenges

and efficiently deploy reliable WSNs. More metrics (connectivity, coverage,

cost, lifetime, packet latency and packet drop rate) are modeled more

practically compared with other works. Especially 3D ray tracing method

is used to model the radio link and sensing signal which are sensitive to the

obstruction of obstacles; network routing is constructed by using widely-used

AODV protocol; the network longevity, packet delay and packet drop rate are

22

1.4. Proposed methodology and work flow: Main contributions

obtained via simulating practical events in WSNet simulator, which to the

best of our knowledge, is the first time that network simulator is involved in

a planning algorithm. Moreover, a multi-objective optimization algorithm is

developed to cater for the characteristics of WSNs. The capability of providing

multiple optimized solutions simultaneously allows users making their own

decisions accordingly, and the results are more comprehensively optimized

compared with other state-of-the-art algorithms.

• iMOST is developed based the novel planning methodology, by integrating

the automatic 3D reconstruction method, the ray-tracing engine and the

planning algorithm, to assist WSN designers efficiently planning reliable

WSNs for different configurations. The abbreviated name iMOST stands

for an Intelligent Multi-objective Optimization Sensor network planning Tool.

iMOST features with: (1) Convenient operation with a user-friendly vision

system; (2) Efficient and automatic 3D database reconstruction and fast 3D

objects design for both indoor and outdoor environments; (3) It provides

multiple multi-objective optimized 3D deployment solutions and allows users

to configure the network properties, hence it can adapt to various WSN

applications; (4) Deployment solutions in the 3D space and the corresponding

evaluated performance are visually presented to users; and (5) The Node

Placement Module of iMOST is available online as well as the source code

of the other two rebuilt heuristics. Therefore WSN designers will be benefit

from this tool on efficiently constructing environment database, practically and

efficiently planning reliable WSNs for both outdoor and indoor applications.

With the open source codes, they are also able to compare their developed

algorithms with ours to contribute to this academic field.

An overview of the planning tool is shown in Fig.1.9. As can be seen, it

contains a user interface and three functionality modules: Image Processing

Module, Ray-tracing Propagation Module and Node Placement Module. They

are developed and embedded together to make the planning tool powerful and

useful, thus contribute on the above mentioned aspects.

The user friendly interface: provides user interaction with the tool and

algorithms. It allows importing objects such as furniture for indoor

23


Glass wall

Brick wall

w1w2

w5

w6w7

w3

D2

D3

D1

D4

w4v1

v2

ray1

TX

ray2 RX

WSN Planning Tool

Image Processing Module Ray-tracing Propagation Module Node Placement Module

Figure 1.9: The proposed methodology and tool.

environment and vegetation and car for outdoor environment; it supports

defining the sensing area AS in the space, pre-deploying nodes at some

specific locations and configuring the parameters such as the transmission

power and receiver sensitivity of the antenna and sensing ability of

sensors; the deployment solution is visually provided to the user by

indicating the locations of nodes, the constructed topology and the

evaluation results.

The Image Processing Module (IPM): is in charge of automatic

3D environment database reconstruction for indoor and outdoor

environments. The algorithm is able to recognize different obstacles

automatically with high accuracy. The birds’ view RGB images of the

outdoor region (taken from satellite camera or sketched from websites

like Google Maps) or the scanned map of indoor space should be

prepared beforehand. Afterwards, the recognized result is vectorized and

the 3D database is built accordingly without human supervision.

The Ray-tracing Propagation Module (RPM): employs the developed

ray-tracing method to trace the propagation path for both radio and

24

1.5. Conclusion

sensing signal. Obstacles are considered along the ray path, therefore

Non-line-of-sight (NLOS) obstruction is detected for the sensing signal

and the multi-path effects and attenuations due to different obstacles are

calculated in a proper way to enhance the accuracy of radio propagation

estimation.

The Node Placement Module (NPM): provides multi-objective

optimized WSN deployment solutions according to the user configuration.

This module employs the proposed multi-objective planning method

and iteratively searches the best topologies to simultaneously trade-off

among the sensing coverage, connectivity, cost, WSN longevity and

packet delivery status.

1.5 Conclusion

In this chapter, the widespread applications of WSNs are introduced in terms of

military, environment, health monitoring and tracking, which nowadays are evolved

to integrate different application categories to construct our world a smart world.

Due to the lack of available experiences and guidance to assist WSN designers

efficiently planning reliable and optimized WSN topology for various applications,

practical and efficient planning methodology and tool are strongly demanded to

facilitate WSN design step so as to reduce human efforts, cost and optimize network

performance before real field deployment.

Several state-of-the-art planning algorithms and tools are investigated and

compared in this chapter, from the computation dimension, optimization objectives,

modelling of objectives and application constraints. Besides of the advantages, the

limitations of them are discussed thoroughly.

A novel planning methodology is proposed in this work, to librate WSN designers

from time consuming and costly 3D outdoor/indoor environment modelling, by

using the proposed automatic image understanding and vectorization algorithm; to

increase the accuracy and efficiency in radio propagation estimation and sensing

signal tracking through the developed ray-tracing engine; to practically model

important metrics; and to comprehensively optimize the performance of WSNs by

25


properly selecting nodes and their properties based on multi-objective optimization

method.

An intelligent multi-objective optimization sensor network planning tool

(iMOST) is developed in this work as well. It is a practical realization of the proposed

planning methodology, the user-friendly interface facilitates the user operations, and

visually demonstrates results. The three functional modules of iMOST work together

to realize the proposed planning methodology. The planning tool, the planning

methodology and the algorithms contribute to this work will be described in details

throughout this dissertation.

26

Chapter 2

3D environment reconstruction

method

2.1 Introduction

WSN designers should be provided with the information of the region where the

WSN will be deployed, so that the proper locations for placing sensor nodes can

be decided to achieve application requirements. The important factors that affect

the deployment are the obstacles and irregular terrains which can attenuate radio

signal strength and block continuity of sensing region; there may exist particular

areas that prohibit to place sensor nodes and also exist some other regions to be

monitored.

A traditional and standard approach to create a 3D model for a region is to build

from scratch using tools such as CAD software or scripts that can be imported to a

demonstrator. Roads, vegetation and building blocks can be described in the form

of primitive 3D shapes based on which terrain data or maps should be provided

in advance: either obtain from manual measurements, from city hall or purchase

from professional company. Therefore, this geometry-based modeling technique is

time-consuming and costly, especially for constructing large-scale scenarios where

many obstacles exist. For many radio propagation experts, the major challenge is

that rather than analyzing the radio performance, they are not skillful at preparing

the environment model. Moreover, according to the survey in the academic

conferences, we notice that some research institutes use the same scenario all the

time to analyse their work without the possibility to test the robustness and accuracy

for different environments.

Chapter 2. 3D environment reconstruction method

2.1.1 3D environment reconstruction from Lidar systems

The 3D environment reconstruction system may consists of automatic building

extraction and reconstruction [62], road extraction and reconstruction [63],

vegetation extraction and reconstruction [64]. Such applications require detailed and

reference models that are still usually created manually. Several recent techniques

[65] aim to increase the level of automation and realism by starting with actual

images of the object or directly digitizing it with airborne laser scanning. Fig. 2.1

and Fig. 2.2 shows the general principle of the airborne laser scanning. The

Figure 2.1: Airborne laser scanning.

Figure 2.2: Example of the DSM and its original image.

standard features of recent airborne lidar systems are their ability to discriminate

between first and last pulse reflections. A laser pulse that is fired over an object

usually has multiple reflections. Some of the laser pulses may be reflected by the

28

2.1. Introduction

top of the object and therefore represents the first returning pulse. The remainder

is likely to be reflected by the ground and hence generates the last returning pulse.

Through computing the arrival time of the laser signal, lidar produces a fast and

highly accurate three-dimensional Digital Surface Model (DSM) [66]. Thereafter the

objects are extracted by using several image processing algorithms such as the edge

detection and shape based classification [67],[68]. The limitation of this method is

the high cost and efforts to obtain the scanned data, as helicopter or airplane is

needed to carry the laser scanner and many human efforts are needed to realize

the task. Therefore it is impractical to use such expensive technology for radio

propagation simulation and network planning issues.

2.1.2 3D environment reconstruction from images

2.1.2.1 Multi-view stereo reconstruction

There are some works proposed to use multi-view stereo reconstructions that are

efficient to provide dense and full 3D reconstructions from multiple views. These

methods can be classified into:

1. Visual hull reconstruction. Obtain the 3D representation of an object through

exploiting the silhouette information. The principle of this method is shown in

Fig. 2.3. This technique assumes the texture of objects can be separated from

the background. Under this assumption, the original image can be converted

into a binary image, which we call a silhouette image. The foreground mask,

known as a silhouette, is the 2D projection of the corresponding 3D foreground

object from a specific view point. Along with the camera parameters, the

silhouette defines a back-projected generalized cone that contains the actual

object. This cone is called a silhouette cone and the intersection of the two

cones is called a visual hull [69] which is a bounding geometry of the actual

3D object.

2. Space carving reconstruction. Generates an initial reconstruction that

envelopes the object to be reconstructed. Then it iteratively removes

unoccupied regions from an explicit volumetric representation. All voxels

falling outside of the projected silhouette cone of a given view are considered

inconsistent and are eliminated from the volume, see Fig. 2.4.

29


Figure 2.3: The intersection of silhouette cones defines an approximate geometricrepresentation of an object called the visual hull. A visual hull has several desirableproperties: It contains the actual object, and it has consistent silhouettes [70].

3. Image-based rendering. Graph-cut algorithm is used in combination of carving

approach, silhouette information to obtain higher precision. Fig. 2.5 shows the

principles of the light field rendering method by [72]. This technique rotate the

source of light along with the object platform, therefore different projections

of the object with constant incident angle of light are taken by fixed camera.

As can be seen, the multi-view stereo reconstruction requires multiple images

of different views for the same scenario, where the objects are a few, in most cases

there is only one object, with a texture quite different than the background. When

the scenario becomes large and complex, or when there is no multiple stereo view

available, this technique is not applicable any more.

2.1.2.2 Single image reconstruction

Most recently, researchers begin to explore reconstruction from monocular view from

a single image [73, 74]. The 3D reconstruction from single image is a challenging and

attractive issue, as unlike human vision system and brain which have been trained for

years and decades to predict the invisible part of an object by observing a single view,

30

2.1. Introduction

Figure 2.4: Example of space carving reconstruction [71]. (a) The plane-sweepalgorithm ensures that voxels are visited in order of visibility with respect to allactive cameras. The current plane and active set of cameras is shown in orange. (b)The shape evolves and new cameras become active as the plane moves through thescene volume.

Figure 2.5: Object and lighting support and the prototype camera gantry [72].

the computer is much simpler, and usually more than most other computer vision

problems, single-view reconstruction is a highly ill-posed problem. To intimate the

human vision, researchers integrate machine learning method to train the dense map

database, to recognize objects or the shapes of specific images. Ashutosh Saxena

et al.[75, 76] patch the images for both 3D location and 3D orientation and use a

Markov Random Field (MRF) to infer a set of ”plane parameters” that belong to

each small homogeneous path in the image. After training the MRF via supervised

learning, the approach (see Fig. 2.6) is able to estimate the relationships between

31


different parts of the image thus generate the depth map for the test images. This

method lacks the ability to model the invisible structure of the objects. In some

works, human interaction is necessary to indicate the important feature markers.

Zou et al. [77] uses a set of auxiliary reference grids to precisely reconstruct both

polyhedral objects and curved-surfaced objects from a single image with unknown

camera parameters. In that method, users should first define the edges and vertices

of the objects, based on which, the camera is calibrated and reference grid is obtained

accurately. Afterwards, the 3D wire frames of object is generated and surfaces are

rendered, as a result it can be applied only when there are a few objects on the image,

otherwise the human efforts spent on marking feature points will be unaffordable.

Original image

Ground truth of

depth map

(a)Training Database (b)Training procedure (c)Test result

Figure 2.6: (a) The database of the depthmap is obtained by using the color mapof laser scan; (b) The feature vector for a superpixel, which includes immediate anddistant neighbors in multiple scales. The relationships are learnt through trainingwith the groundtruth database; (c) An example of the test result which indicate thedepth map of the test image with color scale.

2.2 3D outdoor environment reconstruction

The outdoor environments, where the WSNs will be deployed in our work, are the

European urban cities. The dominant types of obstacles for such environments are

buildings, trees and cars of various colors, textures and shapes.

32

2.2. 3D outdoor environment reconstruction

Although the airborne laser scanning (ALS) can accurately reconstruct large

scale outdoor environment, it requires expensive hardware equipments (e.g.

helicopter, laser scanner), time and a lot of manpower which violate the purpose of

reducing the cost and efforts for deploying WSNs. Whereas the reconstruction from

multi-stereo images are only suitable for simple scenarios with a few objects. It is also

a very tough task to fetch city images with different stereos and views. For example,

Google street view research group record street-level imagery by mounting the street

view camera system on custom road vehicles, trike and modified snowmoble Fig. 2.7.

A lot of human efforts are needed to register the information with the global scale

and to construct the 3D models. Therefore this method is not suitable for our work

either.

In the early 21st century satellite imagery became widely available when

affordable, easy to use software with access to satellite imagery databases was

offered by several companies and organizations. Several other countries have satellite

imaging programs, and a collaborative European effort launched the ERS and

Envisat satellites carrying various sensors. All satellite images produced by NASA

are published by Earth Observatory and are freely available to the public; Satellite

images of different resolutions are available on Google and Yahoo Maps, which

provide us the opportunity to access the satellite images freely. Since all the objects

are visible on the images in pixels, it is convenient to indicate and locate the objects

without measuring the geographic information in the real campaigns, which provides

us the opportunity to reconstruct the cities and outdoor environments conveniently.

However, as the scenario can be very large and complex, it is impractical and

tedious for users manually indicating each primitive on the images. To avoid such

a time and effort consuming procedure, we proposed a solution to automatically

reconstruct the scenarios. This method should be able to simultaneously distinguish

different objects on a image. Moreover, the recognitions must be realized pixel

wisely so that it is suitable for further implementation for radio propagation and

network planning. Thanks to the long historic development of image understanding

researches, there are several state-of-the-art algorithms appeared in the most recent

years which make it possible to recognize multiple objects pixel wisely without any

human supervision.

33


Figure 2.7: Google street view tools. From top to bottom and from left to right:street view car, snowmobile, trekker, trike and trolley.

He et al. [78] and Belongie et al. [79], are examples of those who propose to

identify the objects by shape feature descriptor. Although the detailed shape-based

matching algorithm can be quite different, the principles of them are similar, which

basically consist of 3 steps:

1. Solve the correspondence between shapes;

2. Identify the model class to which the input object belongs;

3. Provide the point correspondences on the matched contours or further refine

the matched points to provide accurate matched result between the actual

shape and reference shape model.

34


The limitation of using shape-based algorithm to understand urban city environment

lays in the variety of shapes for the same type of objects.

Arivazhagan et al. [80] use color and texture features to extract fruit from simple

background. He et al. [81] and Shotton et al. [82] recognize different obstacles based

on machine learning incorporating texture, layout, and context information, which

provide more comprehensive methods to distinguish multiple objects from complex

scenarios. Shotton et al. [82] proposed a texton-boost algorithm that iteratively

selects discriminative texture-layout filters v[r,t](i) to compute weak learners, and

combines them into a strong classifier of the form H(c, i) =∑M

m=1 hmi (c). Each

weak learner hi(c) is a decision stump based on the response

hi(c) =

⎧⎨⎩a[v[r,t](i) > θ] + b if c < C

kc otherwise(2.1)

For those classes that share the feature c ∈ C, the weak learner gives hi(c) ∈{a + b, b} depending on the comparison of feature response to a threshold θ. For

classes not sharing the feature, the constant kc makes sure that unequal numbers of

training examples of each class do not adversely affect the learning procedure. The

machine learning algorithm, to some extent, intimates the learning procedure of

human brain, turns out to be more suitable for our work. Ideally, once the method

learns the features of different objects in different environments, it is able to indicate

pixel wisely each objects on the images robustly no matter how the WSN deployment

environment varies.

2.2.1 Proposed algorithm for outdoor environment reconstruction

In this thesis, the algorithm proposed by Shotton et al. [82] is extended and

sub-sampling and random feature selection techniques are used for iterative learning

of the images. The estimated confidence value of each class for each pixel can be

reinterpreted as a probability distribution using soft max transformation [83] to

calculate the texture layout potentials. The work flow of the proposed method is

indicated in Fig. 2.8.

First of all, a set of images of the target region should be taken: either through

super high resolution satellite camera or through standard hand-held camera. The

35


A large Scenario (a composition of small images of similar sizes)

3D reconstructionRecognition and Segmentation

Data compression

and Vectorization

Figure 2.8: The work flow of the 3D outdoor environment reconstruction.

properties of those images should be similar to that of the training images on the

size, the angel of view, the resolution and so on. The construction of training image

set will be discussed in Section 2.2.2.

Thereafter the developed image understanding algorithm and segmentation

algorithm are applied to automatically recognize objects in the provided images

and extract them pixel-wisely.

Image regularization and vectorization algorithms are employed to regularize

the segmented objects into compact primitive shapes. Their position, rotation and

scale information is registered with reference shapes, so that the feature points can

be extracted and objects are vectorized with a 2D planar coordination. The GIS

information is obtained and transform to the format of longitude and latitude.

At the end, a ’KML’ log file is generated according to the vecotization result.

The file can be demonstrated visually in software like Google Eearth and it provides

an open access of GIS data for all kinds of purpose, especially for radio propagation

simulation and network planning community that have urgent demands on GIS

information of various environments.

2.2.2 Image database

Our method starts from constructing the training of an image database, through

which the features of images are learnt and discriminated for the concerned types of

objects. In this work a bench mark image database and a user defined database are

36


used. The color code for different objects are indicated in Fig. 2.9(c) and example

images of the two databases are shown in Fig. 2.9 and Fig. 2.10 respectively.

The Microsoft Research Cambridge (MSRC) database is composed of 591

photographs of the following 21 object classes: building, grass, tree, cow, sheep, sky,

airplane, water, face, car, bicycle, flower, sign, bird, book, chair, road, cat, dog, body,

and boat. A subset of original images are shown in Fig. 2.9(a). The corresponding

ground truths are shown in Fig. 2.9(b), where each color maps uniquely to an object

class label according to the color codes. In this database all the images have similar

size which is approximately 320× 240 pixels.

Note that, in reality there may exist some scenarios contain other objects rather

than those in the 21-object database and different visual angles than that provided

by the images. To tackle the problem, one can label new object classes manually

by assigning different and unique color codes to create a new database that satisfy

user’s requirements. However, the 21 objects are far many in this work for real world

radio propagation and network planning in the European urban cities where only a

few objects (e.g. buildings and cars) dominant the impacts on radio propagation and

network topology. Besides we intend to use the satellite images, where bird’s view

of the scenario is taken, the perspective of those 21-class images are not suitable for

recognizing objects efficiently. As a result, CEIeurope database is constructed for

recognizing the European style urban city and only four object classes are focused.

The visual angle is the bird’s view from the high resolution satellite camera. A

selection of images and their ground truths are shown in Fig. 2.10 all the images

are 800× 550 pixels with JPEG format and the resolution is the 18th level (among

the 20 levels) on Google Map. The four-object classes are road, building, car and

tree, they are labeled by using the same color code as that in MSRC database.

2.2.3 Image understanding and segmentation algorithm

The meaningful object classes are recognized through a machine learning mechanism,

which consists of a training phase followed by an evaluation phase. In the

training procedure, the Joint-Boost algorithm [82] is employed to compute the weak

learners, which are combined together at the end to compose a strong classifier that

allows multi-object recognition from images. The evaluation procedure recognizes

the objects based on a combination of possibility matrix, which indicates the

37


(a)

(b)

(c)

Figure 2.9: The MSRC labeled image database. (a) A selection of images in the21-class database. (b) The ground truth annotations corresponding to (a),(c) Thecolor codes for the 21 classes.

possibility of how each pixel belongs to each object class. The possibility matrix

is provided by the trained strong classifier from training phase and a sub-cluster

dominant class computation algorithm to refine the unprecise pixels without any

human supervision.

38


(a)

(b)

Figure 2.10: The CEIeurope labeled image database. (a) A selection of images inthe database. (b) The ground truth annotations corresponding to (a).

2.2.3.1 The training phase

All the images, including the training images and the test images, are automatically

converted from RGB color space to Lab color space at the very beginning. Lab color

space with dimension L for lightness and a and b for the color-opponent dimensions,

based on nonlinearly compressed CIE XYZ color space coordinates, is the only way

to communicate different colors across different devices.

The training images are convolved with a 17-dimensional filter bank (see Fig.

2.11) to extract the features of texture. The filterbank consists of: 3 Gaussians,

4 Laplacians of Gaussians (LoG) and 4 first order derivatives of Gaussians. The

three Gaussian kernels (with σ = 1, 2, 4) are applied to each L, a, b channel, thus

producing 9 filter responses. The four LoGs (with σ = 1, 2, 4, 8) are applied to the

L channel only, thus producing 4 filter responses. The four derivatives of Gaussians

are divided into the two x-and y-aligned sets, each with two different values of

σ (σ = 2, 4). Derivatives of Gaussians are also applied to the L channel only, thus

producing 4 final filter responses. Therefore, each pixel in each image has associated

to a 17-dimensional feature vector. The choice of filter-bank is somewhat arbitrary,

as long as it is sufficiently representative. This filter-bank was determined to have

full rank in a singular-value decomposition (see [67]), and therefore there are no

redundant elements.

39


(a) (b) (c)

Figure 2.11: The 17D filter bank. (a) The 3 Gaussians are applied to L,a,b channels.(b) 4 LoGs are applied to L channel. (c) 4 derivatives of Gaussians are for L channel.

�Filter bank

Input image Texton Map

Kd-tree clustering

0

10

20

30

40

50

Figure 2.12: The process of image textonization.

The 17D responses for all training pixels are then whitened (to give zero

mean and unit covariance), and an unsupervised clustering is performed. The

Euclidean-distance K-means clustering algorithm is employed, which can be made

dramatically faster by using the techniques of [84]. Finally, each pixel in each image

is assigned to the nearest cluster center, producing the texton map. The procedure

of image textonization is shown in Fig. 2.12. The texton map is denoted as T where

pixel i has value Ti, i ∈ {1, ...,K}.

40


Afterwards, texture-layout filter is applied to learn the layout of texture and

context information of T . Each texture-layout filter is a pair (r, t) of an image region

r, and a texton index t. Region r is defined in coordinates relative to the pixel i being

classified. For computational efficiency, only rectangular regions are investigated in

our work, though an arbitrary region could be used. A set R of candidate rectangles

are chosen randomly at the beginning of the training procedure to reduce the endless

exploration of the shapes. As illustrated in Fig. 2.13, the feature response of texton

index t at the location i is the proportion of the pixels under the offset region r + i

that have texton index t and it is expresed by (2.2). Outside the image boundary

there is zero contribution to the feature response.

v[r,t](i) =1

area(r)

∑j∈(r+i)

[Tj = t] (2.2)

We define that C is the set of object classes that is within the focus of application.

c ∈ C represents a member of C. N ⊆ C represent a sub-set of the classes.

Joint-Boost algorithm iteratively selects discriminative texture-layout filters v[r,t](i)

to compute weak learners. Each weak learner hi(c) is a decision stump based on the

response

hi(c) =

⎧⎨⎩a[v[r,t](i) > θ] + b if c ∈ N

kc otherwise(2.3)

At each iteration m, the subset N is randomly picked from C. For those classes

c ∈ N that share the feature, the weak learner gives hi(c) ∈ {a + b, b} depending

on the comparison of feature response to a threshold θ. For classes not sharing the

feature, the constant kc makes sure that unequal numbers of training examples of

each class do not adversely affect the learning procedure. At each round, a new

weak learner is chosen by minimizing an error function

Jwse = ΣcΣiwci (z

ci − hmi (c))2 (2.4)

Where each pixel i is attached with a target value zci {−1,+1} (+1 if i belongs to the

class c, and −1 otherwise), wci is the weight function that specify the classification

accuracy of i after m− 1 rounds of learning.

41


Texton Map

t2

1r

2r

t1

(a) Feature responses of v1

(b) Feature responses of v2

Figure 2.13: Calculating the feature responses for textons. (a) the feature responsesfor v[r1, t1]. (b) The feature responses for v[r2, t2].

Theoretically, all the candidates r in R and t in T should be traversed at each

iteration to find the best v[r,t] that produces the minimum Jwse for all the possible

sharing class sets. However, the traversal is expensive and time consuming as there

are |T | × |R| different combination candidates. To efficiently tackle this problem,

each candidate pair of [r, t] is tested with a probability during an iteration and with

thousands of iterations, all the possibilities are likely to be evaluated.

Given the feature filter v and the threshold. The solutions of a+ b, b and kc are

expressed as follows:

b =

∑c∈N

∑iw

ci z

ci [v(i, r, t) ≤ θ]∑

c∈N∑

iwci [v(i, r, t) ≤ θ]

(2.5)

a+ b =

∑c∈N

∑iw

ci z

ci [v(i, r, t) > θ]∑

c∈N∑

iwci [v(i, r, t) > θ]

(2.6)

42


kc =

∑iw

ci z

ci∑

iwci

(2.7)

We assume wci = 1 at the very beginning, and after a new weak learner is selected

at round m, it is updated as

wci := wc

i e−zci h

mi (c) (2.8)

Thus at the end wci = e−zciHi(c) and all the weak learners are combined into a strong

classifier of the form:

H(c, i) =M∑

m=1

hmi (c) (2.9)

The texture-layout filters together with their thresholds and sharing classes

compose the weak learners provided from the training phase, they are eventually

stored in a training database for further evaluation.

2.2.3.2 The evaluation phase

At the evaluation step, all the test images are converted into Lab color format and

convolved with the 17D filter bank, and the responses are clustered to generate

texton maps as the training procedure. The training database generated in the

previous step are applied to the texton maps of all the test images. Hx(c, i) is

computed on each pixel i in a test image x and the estimated confidence value can

be reinterpreted as a probability distribution using soft max transformation [85] to

give the texture layout potentials:

Px(c, i) ∝ exp Hx(c, i) (2.10)

Note that for each image a probability matrix Px is generated with three

dimensions Hx × Wx × |C|, with H and W as the height and width of x, |C| isthe number of classes in C. The temporary recognition result of each pixel i ∈ x,

up to now, is the class with the maximum potential value.

Ct(i, x) = argc

maxPx(c, i) (2.11)

43


An example of the result is demonstrated in Fig. 2.14(b). Although most of the

pixels are classified correctly and different objects are recognized, the precision of

the edges between adjacent objects are not sufficient. Shotton et al. in [82] proposed

a supervised method by manually indicating the misclassified parts, thereafter the

algorithm adjusts those miss-classified parts into the right classes. Note that when

there are a batch of images to be adjusted, the efforts of manually indication for each

misclassified segments will be considerable and time consumption will be very high.

Therefore, we proposed an automatic adjustment method to improve the results

without any human interaction.

(a) Input image (b)preliminary recognition and segmentation

(c) Edge refined by Graph-Cut algorithm

(d) Improved result by subdividing the clusters

Figure 2.14: Multi-object recognition procedure.

2.2.4 Performance enhancement

2.2.4.1 Use of color information for refining edges

Color cues are frequently used in image processing area for edge detection, as the

color gradient variation at the border of adjacent objects are obvious, and that is

also how human eyes distinguish the border of objects. In this work we propose to

utilize the property of color variation information to improve the previous result Ct

efficiently.

First of all, the image is smoothly clustered into K clusters based on the RGB

color distribution. Instead of using normal k-mean cluster method, a graph cut

algorithm [86, 87] is employed to smoothly cut the edges of different clusters.

Because our work requires the labels vary smoothly almost everywhere while

preserving sharp discontinuities that may exist at object boundaries. Fig. 2.15

compares the results obtained from traditional k-mean cluster method to the

Graph-cut algorithm, in both of the results, each pixel is assigned to a cluster and

44


(a) Input images (b) k-mean cluster (c) Graph cut algorithm

Figure 2.15: Comparison clustering results between k-mean cluster and graphcutalgorithm. (a) A selection of the input images from MSRC database. (b) Theclustering result obtained by k-mean cluster.(c) The clustering result obtained bygraphcut algorithm.

labeled with different color code. The cluster boundaries of the graph-cut contain

less noise and are more desired by this work.

Basically the selection of the number of clusters K is arbitrary: the larger the

value is, the more details there will be in the clustering results. As our purpose is

to distinguish the non-trivial color information while maintain a certain tolerance

of variation to avoid the over clustering issue, the proper range of K should be

within 5 to 8 depending on the complexity of the target area. We denote Cdm(k) as

45


the dominant class of cluster k, and it is calculated by selecting the class with the

maximum summation of likelihood for each cluster (see (2.12) and (2.13)),

Px(cj |k) = 1

area(k)

∑i∈k

Px(cj , i) (2.12)

Cdm(k) = argc

maxPx(cj |k) (2.13)

Px(cj |k) is the likelihood that the pixels in cluster k belong to class cj , and

Px(cj , i) is the likelihood matrix of image x calculated by (2.10). Cdm(k) is the

dominant class decision for cluster k. Therefore each object, constraint by color

clusters and class likelyhood matrix, has much clearer boundary than the result

when only Ct is used.

The example results (Fig. 2.14(c)) visually indicate that the boundaries of

adjacent objects are clearer than in Fig. 2.14(b) while the recognition accuracy

is reduced.

Due to the color similarity among different classes, for instances the color of sky

and that of the windshields of cars are similar, parts of buildings and roads have

similar colors, some pixels not in the same classes are labeled with same clusters.

As a result, the dominant class method makes the adjustment inappropriate for the

dominated members.

To tackle this problem, we further divide each cluster into various numbers of

subclusters based on their connectivity property (see Fig. 2.16). The likelihood

matrix Px(cj |ks) is calculated for each subcluster ks by using (2.12). Then Px(cj |ks)is sort in descent order, and the class with the max potential value is the dominant

class Cdm(ks) of ks.

An exception occurs when the class with the second max potential is also the

dominant class of the parent cluster Cdm(k) with ks ∈ k, and if Px(Cdm(k)|ks) ≥μPx(Cdm(ks)|ks), the class of ks is set to Cdm(k) rather than Cdm(ks), otherwise

C(ks) = Cdm(ks). The recognition result after this step is shown in Fig. 2.14(d)

and more results are shown in Fig. 2.17.

46


(a) input image

(b) graph-cut clustering

(c) A selected cluster k (d) sub cluster map ks of k

k

k k1

k2

k

Figure 2.16: Example of sub-clustering based on connectivity property.

2.2.4.2 Image understanding results and comparisons

We perform experiments on a subset of the 21-object MSRC database, and only focus

on 6 out of the 21 classes: building, grass, tree, sky, car and road, which might be

the main objects that affect outdoor radio propagations. The experimental results

are given in Table. 2.1 in a format of confusion matrix.

True class Inferred

class

Building

Grass

Tree

Sky

Car

Roa

d

Building 75.7 0.1 8.1 4.9 9.5 1.7

Grass 2.4 44.3 35.0 0 1.1 17.2

Tree 16.8 0.6 71.6 8.9 1.8 0.3

Sky 10.0 0 5.7 84.3 0 0

Car 12.8 0 6.0 0.4 76.0 4.9

Road 7.1 0 1.5 0 2.6 88.8

Table 2.1: Confusion matrix. Number of cluster in texton booster is 400, the averageaccuracy is 76.1%.

With 1000 rounds of boosting 45% of training images in the database, the number

of texton clusters is set to be 400 to achieve the best average accuracy of 76.1%

whereas the Joint-Boost algorithm gives an overall accuracy of 69.2%. Shotton et

al. [82] then manually moved the misclassified parts to the right classes and the

47


improved result has an accuracy of 72.2%. We notice that the proposed method

by our work outperforms the Joint-Boost algorithm and the edges of adjacent

objects are much more accurate and clearer. Moreover, the proposed method

has the capability to adjust the misclassified pixels to the correct classes: For

instance in the second image of Fig. 2.17, some pixels belonging to the building

are recognized as car by Joint-Boost algorithm. The proposed method is able to

change the decision and move the pixels to the correct classes, and some pixels are

moved to the building class and some are moved to tree class. Hence the proposed

multi-object recognition method is fully automatic and more accurate compared with

the Joint-Boost method, and it can be used to decide the location and the material of

objects. It provides the opportunity to automatically assign the attenuate coefficient

for each object, which will be useful in the following work to increase the efficiency

of radio propagation simulation.

2.2.4.3 Eliminating shadow effect

The available satellite images are taken for different places at different time, and the

solar incident angles are abundantly different. If an object is partially in the shadow

of another object, the discontinuity of texture variation occurs. The shadow effect

has undesired impacts on the image understanding results. Therefore it is important

to detect and eliminate shadow effect.

We employ a scan scheme to scan for each row and column the line segments

that fall inside the shadow region rs, the detection is based on color information.

Although this scheme is arbitrary, it works well for urban city environment where

hardly exists surface with large black color. Before detecting the shadows, a test

image is converted from RGB space to Gray scale space. Each column and row are

scanned separately, and Fig. 2.18 shows the principle of a column scan and row

scan.

The pixels of a column/row are filtered by a 1 pixel× 5 pixel window to realize

zero-phase filtering on the high frequency variations, thus the filtered result (red)

is more sensitive to the significant variations. Thereafter, the line segments, which

begin with a decline trend below the threshold value and end with a climbing trend

above the threshold, are considered belonging to rs. The scanning and detection

procedure is repeated until the entire image is checked. As the histogram varies

48


Input Images Results of TextonBoost Improved Result

Figure 2.17: The comparison of results between the proposed algorithm and [82] ina selection of images.

49


0 100 200 300 400 500 6000

50

100

150

200

250

300

original variation curvefiltered curvestart pointend point

Threshold value

Column Scan

Row Scan (a)

(b)

Figure 2.18: The principle of scanning shadow. (a) Row scan and column scan ofthe shadowed line segments. (b) Filtered curves are calculated and compared withthe threshold. The starting point and end point of each line segment are determinedfor each shadowed region.

50


(a) (b)

Figure 2.19: The result of shadow detection. (a) There exist small holes after thecolumn and row scanning. (b) The final result by filling the holes to reduce noisesin the detection.

with different images, there is no fix threshold value that can be pre-determined for

all the satellite images, hence the value of ϕ should be determined for each image

before the shadow detection. In our work, the mean value of the entire gray-scale

image is considered as the value of threshold ϕ in the plot.

The selected line segments compose grid-style shape with lots of small ”holes”,

which can be seen from Fig. 2.19(a). We eliminate those noises through

morphological operation [88]. Besides depending on the order of provided by

the potential matrix in (2.12), the regions belong to shadow area have very high

possibility to be road class. Assuming that rs is the connected region of the shadow

area. If Cdm(rs) �= road, while the second most possible class is road, the pixels in

rs are assigned to road class.

2.2.4.4 Road detection and regularization

Road detection algorithm is used to enhance the performance of the recognition

algorithm for outdoor environment. If the input image is recognized as urban

city (i.e. through detecting the proportion of buildings, cars), road detection

51


(a) (b)

Figure 2.20: Road segmentation and orientation estimation.(a) The original testimage. (b) The classification result after applying the color information onsegmenting and eliminate the shadow. The pixels of road class are extracted andhighlighted with white color, the background is in black color, and the orientationvector is estimated for each pixel that belongs to road class.

algorithm can be used to regulate the street network of the city and reduce unrealistic

recognition result and improve the accuracy.

Once provided image understanding result from the previous step, the pixels

belong to road class are segmented from other objects with binary value (see Fig.

2.20(b) as an example). The orientation vector is evaluated through R. Wilson’s

orientation estimation method [89]. We assume that all the pixels belong to the

same road are likely to have similar direction vector with smooth change, based on

which, the direction map is grouped into 8 clusters (Fig. 2.21) by using the Euclidean

k-mean cluster algorithm, thus the 360 degree space is divided into 8 sections and

each with a range of 45 degree. Each cluster is then segmented separately and

regions are labeled based on the connectivity as the sub clustering method in the

previous section.

The orientation θe of each connected region is computed by the orientation of the

smallest bounding ellipse (SBE), which is the angle between the x-axis and the major

52


axis of the ellipse, see Fig. 2.22(a) as an example. θe is then compared with the

main direction of this region θm, as indicated in Fig. 2.22(b). If θd = |θe−θm| > θT ,

the centroid orientation Oc(ri) of this region ri is defined as θm to maintain the

continuity with its neighbor clusters, otherwise Oc(ri) = θe.

100 200 300 400 500

50

100

150

200

250

300

350

400

450

500 0

1

2

3

4

5

6

7

8

Figure 2.21: The orientations are clustered into 8 groups

Region expansion is realized to optimize the traffic network, and regularize the

width, direction and connect the isolated road segment. The centroid location Cri is

calculated for each region. Direction of−−−−→CriCrj is represented as O−−−−→

CriCrj

, given the

centroid orientation Oc(ri) and Oc(rj), if |O−−−−→CriCrj

− Oc(ri)| and |O−−−−→CriCrj

− Oc(rj)|are less than an angle threshold ranging from 5 to 10 degree, and the connection

between the two regions does not have important effect to the recognized buildings,

the two regions are supposed to be connected. This assumption is useful when only

a small proportion of pixels belong to the road class are misclassified, and they, by

mistake, disconnect ri and rj that are assumed to be in the same road segment.

53


e�

m�T�

(a) (b)

Figure 2.22: A toy example of the comparison between θe of the SBE and the mainorientation θm. (a) The SBE of a region where the orientations belong to the samedirection cluster. (b) Computation of the difference between θe and θm.

Two regions are expanded and connected by rectangles once the aforementioned

constraints are fulfilled. The rectangle is expressed by

r = [O,C,W,L]

where O = O−−−−→CriCrj

is the orientation of the rectangle, C is the central of the

rectangle, W is the width that equals the median distance from the boundary to the

skeleton and L is the length which equals the distance between furthest endpoints

of the two regions. The pixels locate within the rectangle are assigned to road class.

The pseudo code for the aforementioned road detection and expansion algorithm

is:

Listing 2.1: Road detection and region expansion

Segment road class from image x;

compute the orientation map Mo;

k-means cluster with 8 sections(45◦);

for each connected region ri

{

compute Oc(ri)

}

54


for each region pair (ri, rj) and i �= j

{

if (orientation is similar)

{

WR = roadwidth(ri, rj)

HR = roadlength(ri, rj)

OR = O−−−−−→Cri

Crj

CR = median(−−−−→CriCrj )

rectangle: R(CR, OR,WR, HR)

if (Σi∈R[C(i) /∈ road] < 0.5Area(R))

∀i ∈ R, C(i) = road

}

}

Finally, the result of the road detection algorithm is shown in Fig. 2.23.

Compared with the result without any enhancement, the accuracy is significantly

improved after applying the color information and road detection algorithm.

(a) (b) (c)

Figure 2.23: The result of road detection. (a) The test image. (b) The result withoutany enhancement. (c) The result of road detection.

2.2.5 Shape matching and vectorization

Real world objects can be expressed by geometric primitives in the computer graphic

area and CAD systems. In constructive solid geometry, primitives are simple

geometric shapes such as a cube, cylinder, sphere, cone, pyramid, torus for 3D

applications. Whereas a polygon is traditionally a plane figure that is bounded by a

55


closed path, composed of a finite sequence of straight line segments (i.e. by a closed

polygonal chain), which is useful in 2D applications.

The objects recognized by our image understanding algorithms, are to be

described as 3D geometric primitives as well. By means of recording the closed vertex

path for each object, the digital environment database is constructed. Even though

the boundaries of recognized objects can be directly used to compose polygons, they

contain far more number of pixels than that is actually needed, and the information

stored in database will be highly redundant by doing so. A logical and natural

way to tackle this problem, is to extract the feature points from the boundaries of

segmented objects. As the proposed image understanding algorithm is not able to

obtain 100% accuracy, there are still errors and coarse edges at the boundaries which

make the detection of feature points a big challenge for some very simple primitives.

Further more, in urban cities and indoor environments, an overwhelming number

of objects have regular planar shapes such as quadrilateral, triangle, circle that can

be described conveniently with only a few vertexes. Therefore we develop a shape

matching algorithm to detect the shape for all the recognized objects and register

them with the corresponding corners by rotation and deformation.

In this section, the methodology for shape matching and registration to vectorize

the recognized objects and compact database is introduced. At the end, the 3D map

is written to KML format log file to be demonstrated visually on software (e.g. on

Google Earth) and be accessed by other people.

The facades of buildings in European urban cities are often seen with the shapes

shown in Fig. 2.24, the set of such shapes is named as S in this work.

Basically, shape-based image retrieval consists of measuring the similarity

between shapes represented by their features. Some simple geometric features can be

used to describe shapes. Usually, the simple geometric features can only discriminate

shapes with large differences. Therefore they are usually used as filters to eliminate

false hits or combined with other descriptors and they are not suitable to be used

as stand-alone shape descriptors. A shape can be described from different aspects

including: Rectangularity, Solidity, Circularity ratio and Zernike moments (ZM).

56


Figure 2.24: Frequently seen facade shapes of buildings. From top to bottom andfrom left to right are: circular, triangular, rectangular, H, L, Cross and G shape

2.2.5.1 Rectangularity

Rectangularity represents how rectangular a shape is, i.e. How much it fills its

minimum bounding rectangle:

Rectangularity = AS/AR (2.14)

where AS is the area of a shape; AR is the area of the minimum bounding rectangle

box.

2.2.5.2 Solidity

Solidity describes the extent to which the shape is convex or concave and it is defined

by

Solidity = AS/H (2.15)

where H is the convex hull area of the shape. The solidity of a convex shape is

always 1.

57


2.2.5.3 Circularity ratio

Circularity ratio represents how a shape is similar to a circle. It is the ratio of the

area of a shape to the area of a circle having the same perimeter:

Circularity = AS/AC (2.16)

where AC is the area of the circle having the same perimeter as the shape. Assume

the perimeter of the region is P, AC = P2/4π. As 4π is a constant, the Circularity

can be rewritten as

Circularity ∝ AS/P2

2.2.5.4 Zernike moments

Zernike Moments (ZM) are orthogonal moments [45]. The complex Zernike moments

are derived from orthogonal Zernike polynomials:

Vnm(x, y) = Vnm(r cos θ, sin θ) = Rnm(r)exp(jmθ)

where Rnm(r) is the orthogonal radial polynomial:

Rnm(r) =

(n−|m|)/2∑s=0

(−1)s(n− s)!

s!× (n−2s+|m|2 )!(n−2s−|m|

2 )!rn−2s

Zernike polynomials are a complete set of complex valued functions orthogonal over

the unit disk. The Zernike moment of order n with repetition m of shape region is

given by:

Znm =n+ 1

π

∑r

∑θ

rs(r cos θ, r sin θ) ·Rnm(r) · exp (jmθ) r ≤ 1 (2.17)

2.2.5.5 Shape matching and registration

As mentioned before, there is no a single shape descriptor that is able to tell so

many shape types, as the tuning of shape parameters is a challenging problem. In

this work, an hierarchical matching heuristic is employed by swapping the objects

into the proper branch at each level to hierarchically detailing the shape description.

58


As a matter of course, different discrimination metrics are applied at different levels

and the state flow is shown in Fig. 2.25.

oD

rD

Figure 2.25: Hierarchical shape matching.

At the root of the hierarchical model, the discriminator Dr is in charge of

distinguishing circle, triangle and the rest of shapes in S. Then at the second level,

the 3rd branch for the rest shapes, discriminator Do is trained for the remaining 4

shapes. By using the Rectangularity, Solidity, Circularity ratio and Zernike moments

(ZM) (Eq. (2.14) to Eq. (2.17)) of the regions, the discriminator is trained in Neural

Network to learn the shape features. The training shape set is built by rotating,

transforming and deforming the shapes shown in Fig. 2.24. Each training sample

is black-white image of size 255 × 255, where the shape is in black color and with

variation on orientations, scales. Dr is obtained by learning circles as type Sc,

triangles as type St and the remaining types as So; Do is obtained by learning

rectangle as type Sr, H shape as type Sh, cross shape as Scro and the G shape as

SG.

Dr and Do are trained separately by using Bayesian regulation backpropagation

function. The results of the shape matching are listed in table 2.2 and shown in Fig.

2.26.

The accuracy is 100% without adding noise at the boundary of the test shapes,

and 90% when adding the noise for applying to the recognized results where there

are no serious errors. All the typical shapes in S are stored in data structure that

59


Table 2.2: Shape matching result

��OutputInput

Cicle Triangle Rectangles H L Cross G

Circle � � � � � � �

Triangle � � � � � � �

Others � � � � � � �

Rectangles � � � � � � �

H � � � � � � �

L � � � � � � �

Cross � � � � � � �

G � � � � � � �

contains the features like the centroid, orientation, scale of sides and so on, which

also depend on how those shapes described in geometry.

Once provided with the shape matched result. The objects are registered

with the corresponding shape features to search the most suitable rotation and

deformation by minimizing the evaluation metric. Assume S′ is the candidate shape,

AS′ =∑

i∈x[i ∈ S′] represents the area, and AIS′ =∑

i∈x[i ∈ S′ && i ∈ S] represents

the actual area of pixels that belong to S′ and S. The metric that evaluates

the candidate shape is expressed by (2.18). The formula sums the ratio between

candidate shape area and actual union region and the ratio between original shape

area and the union region, which indicates how well the candidate shape matches

with the original region.

E(S′) =AS′

AIS′+

AS

AIS′(2.18)

The object becomes finding the S′b that provides minimum E

S′b = argS′

minE(S′) (2.19)

Thereafter, a iterative procedure is imposed on the shape matching and

recognition through Eq. 2.18 and Eq. 2.19.

2.2.5.6 Delineating the buildings

Recalling that the centroid of ri is Cri : {x, y} and the orientation of ri is Oc(ri). If

ri ∈ Sr, the polygon can be described by four vertexes (A, B, C and D) that are

60


H shape

Rectangle

L shape

Cross

U shape

Figure 2.26: Successful result of shape matching. Each shape of Do is correctlymatched.

constraint by the following conditions:

C =

⎧⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎩

yI′−yAxI′−xA

= tan θ,yI′−yCrixI′−xCri

= − 1tan θ

−−→AD = 2

−→AI ′

−−→DB = 2

−−−→DCri

−→AC = 2

−−−→ACri

(2.20)

Where I ′ is the median point of−−→AD and θ = Oc(ri) is the orientation of the shape.

Note that the recognized region r is not perfect and we need to match the shape

61


with the condition in Eq. 2.20 by adapting θ and side lengths with coefficients:

C′ =

⎧⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎩

yI′−yAxI′−xA

= tan θ′,yI′−yCrixI′−xCri

= − 1tan θ′

−−→AD′ = α

−−→AD

−−→AB′ = β

−−→AB

−−→DC ′ = β

−−→DC

(2.21)

Where θ′ ∈ [θ − φ, θ + φ], normally φ ≤ 45o, and α(or β) ∈ [1 − τ, 1 − τ ] where τ

ranges from 0 to 0.5. Fig. 2.27 shows the procedure of how a rectangle is registered

by rotation and deformation.

100 200 300 400 500

100

200

300

400

500

600

Rotation

Deformation

Rotation

A B

C

D

I’ Cr

Figure 2.27: An example of registering the rectangle primitives.

62

The similar heuristic can be applied to H shape, L shape and G shape by finding

the vertexes that satisfy the conditions. Circular buildings are tested by a set of

circles that centered at Cri and radius R ranging from the Bmin to Bmax which

represent the boundary points that have the minimum and maximum distances

with Cri .

2.2.5.7 Writing geometric primitives to KML file

All the vectorized information are written into Keyhole Markup Language (KML)

format, which is an XML notation for expressing geographic annotation and

visualization within Internet-based, two-dimensional maps and three-dimensional

earth browsers. Google Earth was the first program able to view and graphically

edit KML files. Other projects such as Marble have also started to develop KML

support. The advantage of using KML format is that the format is simple and

descriptions are convenient to create for polygons, points and lines (see Listing 2.2

as an example of the script). It provides an open access possibility to the public and

can be demonstrated in many popular free earth browsers as shown in Fig. 2.28.

Listing 2.2: Common used Definitions in KML file

<?xml version="1.0" encoding="UTF-8"?>

<kml xmlns="http://www.opengis.net/kml/2.2">

<Style id="xxxx">

<LineStyle> </LineStyle>

<PolyStyle> </PolyStyle>

</Style>

<Placemark>

<Point>

<coordinates>longitude, latitude, altitude</coordinates>

</Point>

<Polygon>

<coordinates>

longitude1, latitude1, altitude1

longitude2, latitude2, altitude2

...

</coordinates>

63


</Polygon>

</Placemark>

</kml>

Figure 2.28: KML shape description of the reconstructed result demonstrated onGoogle Earth.

2.3 Indoor environment reconstruction

Unlike the outdoor environment reconstruction, where the objects locate apart

from each other and the scale is large, the indoor environment reconstruction

is much easier. A method is developed to automatically model the 3D indoor

environment from a scanned 2D map if the digital map is not available from the

administrator. The reconstruction method consists of four steps: image calibration

and classification step which recognizes walls from the scanned map; thinning and

feature points extraction to compress the edge information; edge smoothing and

vectorizing to build the 3D database. The reconstructed 3D scene is demonstrated

to users at the end.

64

2.3. Indoor environment reconstruction

(a)

(b)

Horizontal

Vertical

Distortion

Figure 2.29: Image calibration and segmentation. (a) Scanned map. (b) Calibratedand segmented result.

2.3.1 Image calibration and classification

Walls in the maps are normally labeled with black lines as shown in Fig. 2.29(a).

Once the map is scanned, the image is converted from RGB color to gray color.

Because the floor plan on paper has some distortion and transformation during

scanning, the image should be calibrated with 2D horizontal-vertical direction to

make sure that the axis of map matches well with real world axis. Thereafter, each

pixel in the image is checked and pixels that belong to walls are recognized by using

a single classifier h(i),

h(i) =

⎧⎨⎩1 p(i) < v

0 p(i) ≥ v(2.22)

where p(i) is the color of ith pixel in the image, v ∈ [0, 255] is the color threshold.

If the pixel value is less than v, pixel i is marked by 1 which indicates the wall,

otherwise it is marked with 0 which is the background. Hence, a segmentation

result is obtained after this step, see Fig. 2.29(b).

65


2.3.2 Thinning and feature points extraction

In 2D vision of indoor scene, a wall can be expressed by an edge E < V1, V2 > with

V1 and V2 as two endpoints. As can be seen in Fig. 2.30, the extracted pixels from

the previous step contain redundant information for constructing walls. Thus they

are thinned to lines with 1-pixel width based on conventional thinning method and

deliver the skeleton of wall efficiently.

--Extracted pixels--Thinned line--Feature point--non-critical point

Figure 2.30: Thinning step and feature point extraction

However, as can be seen, the thinned lines (skeletons) still contain a large number

of pixels which are not necessary to describe the critical variation of edges and some

pixels are noises with high level fluctuations in a very short distance. Therefore,

those non-critical points should be eliminated and only critical points along the

thinned lines should be maintained. The critical points are also called feature points,

which are either conjunctions shared by edges of different directions or the endpoints

of edges. As a result, they can be used to represent the endpoints of edges. In this

step, Harris-corner algorithm [90] is employed to search critical points throughout

the segmented pixels and the results are demonstrated in Fig. 2.30 and Fig. 2.31.

2.3.3 Smoothing and vectorizing

Critical points are clustered and map is vectorized in this step. Four windows

with different directions are used to cluster feature points that belong to the same

line. Fig. 2.31 shows the shapes of windows and an example of smoothing and

regularization. As can be seen, the green dots are critical points obtained from

previous step. After applying the windows, the locations of points are adjusted

and marked by blue dots. Based on the smoothing algorithm, two points belong

66

2.3. Indoor environment reconstruction

Windows with different directions

Smoothed points Endpoints for regularizing edgesCritical points

Figure 2.31: Smoothing and regularization.

to the same cluster and with largest distance are selected as endpoints of an edge.

This step reduces the number of unnecessary planes and the database information

is therefore compressed.

3D scene consists of obstacles which have shape description and material

information that describe how signal strength is attenuated. Different materials have

different attenuation values. The vectorization results are stored in ’.txt’ file, and the

format is shown at the right side of Fig. 2.32. Vertices in 3D format are expressed

as Vi = [xi, heighti, yi]. In this work, all the walls are assumed to be quadrangle and

have 4 vertices as indicated in Fig. 2.33. However, all the edges are discovered in 2D

vision and only V1 = [x1, 0, y1] and V2 = [x2, 0, y2] are known. Therefore, once the

height of wall is known, V3 = [x1, height, y1] and V4 = [x2, height, y2] are generated

by assigning value to height. The structure of wall can be expressed by

plane =

{V1 V3 V2 V4

material

}′

where material is the index of material such as glass, wood, brick, concrete and so

on.

67


NUMBER OF PLANES 880 0 16.7 10 3 16.757.7 3 16.757.7 0 16.727.8 0 12.8 127.8 3 12.830.9 3 12.830.9 0 12.8… … …

Mismatched edges

Figure 2.32: Vectorization result.

Figure 2.33: A wall is described by four vertexes.

2.3.4 Demonstration and analysis

The proposed 3D indoor reconstruction method is programmed in MATLAB, and

run on a PC equipped with Intel i5-760 CPU of 2.8 GHz frequency. Fig. 2.32

compares the computed edges with the map. Only two edges have relative big

errors (0.35 m) and one wall is missed classified, the rest edges match well with

the map with an average error of 0.13 m. In this example, 86 edges are extracted.

As the heights of walls are identically 3 m, they are included automatically to the

result. Besides, after the endpoints are fetched, the horizontal range (X − range)

and vertical range (Y − range) of target space are also obtained, therefore floor

and ceiling are added automatically to the vectorization result and 88 planes are

constructed at the end.

The 3D views of the reconstructed indoor environment are demonstrated in Fig.

2.34. The size of the target space is 57.7 m×3 m×16.7 m. The time for constructing

68

2.4. Conclusion

Figure 2.34: Reconstructed 3D indoor map in different views.

this indoor map is 5.8 s, compared with conventional method which took almost 1

day per person by manually typing 88× 4 endpoints.

Fig. 2.35 is another example of 3D indoor reconstruction by randomly searching

floor plan on Internet. We select the East Lansing map provided on the web

page [91]. The size of the region is 77.4 m × 3 m × 36.6 m, the proposed

indoor reconstruction method took place on the downloaded map, it took 23.5 s

to reconstruct 156 planes on the floor plan. The recognized edges match well with

the original image (see Fig. 2.35(b)) and the 3D reconstruction is viewed in Fig.

2.35(c). As a result the proposed method significantly reduces time and human

efforts on modeling different indoor environments.

2.4 Conclusion

3D environment modelling is a difficult issue for radio engineers and network

designers who do not have expert knowledge on this topic. In conventional

approaches of constructing a 3D environment database, either a large amount of

money should be spent to buy the digital vector map from professional companies,

or time and human efforts are needed to manually measure and reconstruct the

real campaigns. The expenses over this issue constraint the researchers and private

users on validating their proposals or models, and they can be analyzed on a limited

number of scenarios that are freely accessed by public such as the Munich scenario

measured by COST 231 group [92]. To the best of our knowledge, this is the first time

69


(a) Original map of East Lansing

(b) Vectorizing

(c) 3D view of the reconstructed result

Figure 2.35: (a) A toy example by using the map downloaded from [91]. (b) Thevectorization result and the 3D view of reconstruction (c).

70

2.4. Conclusion

that the advantages of image understanding algorithm is applied to automatically

reconstruct 3D outdoor and indoor scenarios for signal propagation and network

planning purpose. The principle of reconstruction method is introduced in this

chapter and the proposed algorithms are detailed as well. The experimental results

on outdoor reconstruction indicate that the algorithm is able to accurately recognize

different objects from the satellite images of the target regions, it offers a flexibility

to reconstruct different types of environment by training different environment

database. The indoor reconstruction algorithm is proposed to reconstruct the indoor

database from scanned floor plan, the efficiency and accuracy is satisfied without

any human interaction. Thereby, the 3D environment reconstruction methodology

proposed in this work significantly reduces human efforts, cost and time spent on

reconstructing a 3D geographic database.

71

Chapter 3

Ray-tracing engine and radio

propagation modelling

There are numerous algorithms and methods that are developed to optimize the

deployment, routing protocols, power consumption plans and etc. for WSNs. Before

real implementation, those algorithms are evaluated through simulations which

generally take place in network simulators such as NS-2, TOSSIM, EmStar or

OMNeT++. Kamarudin et al. [93] review a variety of realistic propagation models

for WSNs and discuss the modeling of vegetation propagation model in OMNeT++

simulation platform. They prove that propagation model has strong impact on the

evaluation of network performance. However, the aforementioned simulators employ

propagation models as simple as Free-space model and Log-normal model, which are

too optimistic and independent of environment. The evaluations of those protocols

and algorithms based on those empirical propagation models are not rigorous, which

will lead to inappropriate algorithm design. Therefore accurate radio frequency (RF)

propagation modelling and simulation are very important at the pre-deployment

phase of WSN for predicting the design performance.

There are many RF propagation simulation works focused on mobile

communication. RF measurements have been made in several cities such as Munich

scenario [92] and Ilmenau scenario [94]. Based on those experimental results, other

works are done to model the radio propagation, tune critical parameters and evaluate

coverage performance of base stations. Furthermore, signal attenuation parameters

are also tested when signal penetrates through obstacles of different materials and

sizes, such as different thickness of bricks, concrete, metal, etc.

Most of the miniaturized wireless sensor devices implement ZigBee or 802.15.4

communication protocols, while there are only a few research works focused on

modeling RF propagation for ZigBee WSNs, and even fewer work on measuring

Chapter 3. Ray-tracing engine and radio propagation modelling

ZigBee radio propagation in reality. For this reason, recently more researchers

focus on modeling radio propagation for WSN and several experiments are made to

characterize ZigBee propagation features in both indoor and outdoor environments

[95], [96]. Korkalainen et al. [38] discuss advanced mobile signal tools that

might be suitable for estimating 2.4 GHz ZigBee protocol. In order to ensure the

results accurate and the simulation efficient, the tools should have features like

3D calculation, accurate environment modeling and radio modeling. In [97], the

authors characterized wireless channels for indoor propagation at 2.4 GHz, but only

direct path is considered. They conclude that free space propagation is unreliable,

log-normal model is not completely matched with the trend of curvature of real

measurements and the multi-wall-floor model is the most reliable and accurate

among the discussed models. However, only considering the direct path is unilateral,

as diffractions and reflections are also vital.

Generally speaking, ray tracing algorithm has very high accuracy [98, 99, 100]

compared with the empirical propagation models. It is obstacle sensitive which

is able to trace multi-path effects in real world signal propagation. While the

computation load is also very high as the method tests every intersection along

the ray path. Especially when scenario becomes large, traditional method might

take hours or days to finish simulation. Therefore many algorithms are developed

to overcome the aforementioned drawback by slightly reducing accuracy as the

compensation.

Beam tracing algorithm [101] extends ray tracing algorithm to reduce intersection

tests, as well as overcome sampling problems. Dominant path tracing algorithm [102]

is developed to avoid redundant calculation, because the authors believe that 98%

of received power is contributed by only a few radio rays. Ray tube tree method

[103] increases the preprocessing speed in constructing trees for ray-tracing.

A 2.5D outdoor ray launching tool is presented in [104]. The tool is very fast but

the resolution is low (7 m) and the computation load is decreased by reducing 3D

rays to 3D Line of sight (LOS), 2D Horizontal diffraction and reflection (HDV) and

modified 2D Vertical diffraction (VD), therefore the accuracy of simulation result

is constraint by the reduced resolution and the types of rays. A 3D ray-optical

approach is presented in [105]. The calculation is in 3D, however in order to

accelerate the calculation, the method preprocesses the environment by dividing

74

3.1. Space division

the obstacles into tiles and the edges into segments, and ray paths are limited to

only search over the tiles and segments.

As a result the time consumptions of existing ray tracing methods are decreased

by either reducing dimension of rays or reducing the details of the environment,

the simulations are efficient while inaccurate. Therefore, in order to make the RF

simulation method advanced, efficient efficient 3D ray tracing and accurate radio

modelling should be developed and integrated. In this chapter, a 3D ray tracing

engine is developed based on space division and polar sweeping is developed to

efficiently search rays and the modelling of radio propagation is proposed to consider

the obstacles and their order along the signal path. Our work is benefited from this

part in terms of:

1. High accuracy in radio estimation for both indoor and outdoor environment.

2. High efficiency which allows the planning algorithm employing such technique

to practically model the radio connectivity and sensing coverage with

acceptable computational speed.

3. Generality. It can be used for the automatically reconstructed environment

database from previous chapter as well as other GIS databases provided by

designers.

3.1 Space division

Before tracing rays, a 3D environment model, either from the aforementioned

automatic reconstruction method or from other modelling methods, is loaded to

the ray tracing engine, and the target space is split by using kd-tree algorithm.

The target space is divided into small cubes with different volumes to balance the

number of obstacles among the cubes. Vlastimil Havran did an extensive study of

available spatial subdivision schemes (including regular grids, nested grids, octrees

and kd-trees) and concluded in his thesis [106] that the kd-tree beats the other

schemes in most cases. And Ingo Wald claimed that it is possible to limit the

number of ray-triangle intersections to three or less using a well-constructed kd-tree

[107].

75


A kd-tree is an axis-aligned Binary Space Partitioning (BSP) tree. Space is

partitioned by splitting it in two halves, and the two halves are processed recursively

until no half-space contains more than a preset number of primitives. While kd-trees

may look like octrees at first, they actually are quite different: An octree always

splits space in eight equal blocks, while the kd-tree splits space in two halves at a

time. The most important difference though is that the position of the splitting

plane is not fixed. In fact, positioning it well is what differentiates a good kd-tree

from a bad one. Consider the images in Fig. 3.1:

(a) (b) (c)

Figure 3.1: Space division by kd-tree. (a) A scene with some primitives. (b) A badexample of kd-tree division. (c) A well constructed kd-tree.

Fig. 3.1(a) shows a scene with some objects in it. In Fig. 3.1(b), this scene is

subdivided; the first split is the vertical line, then both sides are split again by the

horizontal lines. The split plane position is chosen in such a way that the number of

primitives on both sides of the split plane should be roughly the same. Although this

may sound like a good idea at the first glance, it actually isn’t: Imagine that a ray

traverses through this subdivided scene but it never has chance to pass through an

empty voxel. If we keep adding planes to this kd-tree following the same rule as Fig.

3.1(b), the result will end up with a tree that does not contain a single empty node,

which is the worst possible situation, as the ray tracer has to check all primitives

in each voxel it travels through. The subdivision in Fig. 3.1(c) is well constructed:

A different heuristic by Jacco Bikker [108] is employed to determine the position of

the split plane. The algorithm tries to isolate geometry from empty space, so that

76

3.2. Polar sweeping

rays can travel freely without expensive intersection tests and therefore our space

division mechanism employs the kd-tree algorithm proposed in [108].

3.2 Polar sweeping

At the beginning of ray tracing, the target environment is polar swept. Conventional

polar sweep method is well known to solve geometric problems, by sweeping a line

across the plane and halting at points where the line makes the intersection with

any object within the target region. As depicted in Fig. 3.2, the solution is partially

computed at those intersection points, so that at the end of sweep, a final intersection

result is available.

Polar sweep

Figure 3.2: Conventional Polar sweep.

In this work, conventional polar sweep algorithm is modified by bending the

direction of the line whenever intersection occurs. A 3D line is rotated clock wisely

and bottom up centering at the transmitter (TX) to discover for each direction the

first intersected point and its corresponding plane. The rule of reflection is then

applied to the line, which bends original direction of the line. Intersected planes

are recorded whenever the direction changes at halting points. This procedure is

repeated for each candidate direction, and each 3D line terminates shooting after

77


TX

Glass wall

Brick wall

w1

w2

w5

w6w7

ray1

ray2

ray3

w3

D2 D3

D1

D4

w4

Figure 3.3: Polar sweep.

a maximum number of intersections is reached or when the boundary is touched.

Therefore, by sweeping the entire scene, all the possible orders of reflection planes

and diffraction cones are discovered. The order of reflected plane is stored as a

matrix with dimension of N ∗depth, where N is number of possible reflection paths,

depth is the maximum depth of reflections predefined in the ray tracing engine. For

instance, the depth = 4 in Fig. 3.3 indicates no more than 4 reflections in a ray

path, and the reflected plane is expressed as Ref plane. When visible plane exists

and the length is less than depth, the ID of intersected primitive plane is recorded

in a order along the signal path, such as w1 → w2 → w5 → w6. When the number

of planes in a path is less than depth, NULL is assigned to the remaining elements:

Ref plane =

⎡⎢⎢⎢⎢⎢⎣

w7 NULL NULL NULL

w1 w2 w5 w6

w5 w3 NULL NULL...

......

...

⎤⎥⎥⎥⎥⎥⎦ (3.1)

Diffraction happens at the convex edges, the diffraction points are extracted

based on the resolution of z direction and Dif cone is a structure used to store all

the information of a diffraction cone.

structure Dif_cone{

point;

78

3.2. Polar sweeping

visible_plane;

invisible_plane;

}

Where Dif cone.point indicates the location of the diffraction point,

Dif cone.visible plane is the ID of visible plane it incidents with, and

Dif cone.invisible plane is the invisible plane connected with the visible plane, if

there does not exist such a plane, Dif cone.invisible plane = NULL.

In ray tracing method, there exist many ways to search reflection paths. In this

ray tracing engine, ray launching and ray tracing method are combined in order

to be efficient. Ray launching is executed in the polar sweep step. Ray tracing

happens when knowing the location of receiver (RX). After polar sweeping, the

rays are traced for each RX, direct path, reflection paths and diffraction paths are

selected and the received power strengths (RSS) are calculated separately according

to models given in the following subsections.

3.2.1 Direct path

A direct ray is launched from TX to RX, and each intersection is recorded when

the ray penetrates through objects. The radio propagation model (3.2) which is

developed by us in [109], is used to estimate power loss of direct path between TX

and RX.

Lp = 10 n log10d+ Lobstacle (3.2)

Where n is path loss coefficient ranging from 2 to 5 and the value is dependent on

the propagation environment, i.e. in free space n = 2, in others such as urban or

rural environments 2 < n < 5. d is the direct 3D distance between TX and RX.

Lobstacle is the power loss due to obstacles encountered along signal path.

Lobstacle =

N∑i=1

l(i)α(i)(i−1) (3.3)

It is obtained by accumulating power reduction of each obstacle along the signal

path. l(i) is attenuation parameter of the ith obstacle, α(i) ranges from 0 to 1,

which is penetration rate of the material of ith obstacle. α(i)(i−1) decreases when i

79


RX1

RX2

TX

Glass wall

Brick wall

Figure 3.4: Direct path.

increases, meaning that the first object, with which the signal intersects, produces

the most significant power loss.

Fig. 3.4 gives an example of calculating (3.2) for direct path. There are 2 RXs:

RX1 and RX2, and one transmitter TX. Signal penetrates through two glass walls

before reaching RX1, whereas another signal path penetrates through one brick wall

to reach RX2. For RX1, l(1) = 2 dB, l(2) = 2 dB, assuming α = 0.9 as a constant

value, we have Lobstacle = 2 + 2 × 0.9 = 3.8 dB. For RX2, l(1) = 10 dB, we have

Lobstacle = 10 dB. Eq. 3.2 indicates that not only distance, but also the materials

of obstacles and their orders have impacts on received signal strength.

In the ray tracing engine, the electric fields of radio wave are computed instead

of calculating the path loss in dB, as the phase of propagation wave significantly

affects the final result, thus resulting:

Edir(d) =Et

d√n∏N

i=1 Lobstacle(E)(i)e−j2πkd

(3.4)

Where Et is the transmitted electric field at the sender. Lobstacle(E)(i) is

converted from Lobstacle(i), k is the wave length of radio, which is C/f , C is the

speed of light and f is the frequency in Hz.

80

3.2. Polar sweeping

Glass wall

Brick wall

w1

w2

w5

w6w7

w3

D2 D3

D1

D4

w4

v1

v2

ray1

TX

ray2 RX

Figure 3.5: Reflection path searching.

3.2.2 Reflection path

Starting from RX, each possible path discovered in Ref plane is traversed. The

procedure is shown in Fig. 3.5, only two rays are shown as example. According to the

second possible path in Fig. 3.3 and Eq. 3.1, Ref P lane(2) = [w1 w2 w5 w6], virtual

source v1 of w1 is generated by mirroring TX along w1, v2 is obtained by mirroring

v1 along w2. There are two reflection paths in this plane order, one is through

TX → w1(v1) → RX, and the other one is TX → w1(v1) → w2(v2) → RX. w5

and w6 are ignored for current order. As a result, the format of reflection path

matrix is expressed as:

Ref path =

{w1 v1

w1 v1 w2 v2(3.5)

Note that although the example plot is by projecting 3D space into 2D, all the rays

are traced in 3D approach. With this searching method, all candidate 3D paths

from TX to RX are recorded at the end, and RSSs due to reflections are computed

according to the model in Uniform Theory of Diffraction (UTD) [110]. For a RX at

point P , its electric fiel Eref reflected from point Q is expressed by:

Eref (dQP ) = Ei(Q)Re−j2πkdQP (3.6)

81


TX

Glass wall

Brick wall

w1

w2

w5

w6w7

D1

D2 D3

D4

w3

w4

RX

Figure 3.6: Diffraction path searching.

Where Ei(Q) is the field of ray incident at Q. R is the reflection coefficient related

to the material of encountered plane.

3.2.3 Diffraction path

Diffractions happen at the diffraction points found in Dif point. When RX is

shadowed by planes, diffraction will play a relative important role in RSS. Fig. 3.6

shows that a RX is in the diffraction region of diffraction point D2. The diffraction

field is calculated by (3.7),

Edif (dQP ) = Ei(Q)De−j2πkdQP (3.7)

where Q represents D2 and P stands for RX in this case, D is diffraction coefficient,

D =(1 + cos (θ))

2.0√2πf

,

and θ = |(∠(TX,RX)− ∠(TX,Q))| is the angle differential between the rays from

TX to Q and from TX to RX.

At the end, according UTD theory, the Etot at P is the combination of Edir,

Eref and Edif .

Etot = Edir + Eref + Edif (3.8)

82

3.3. Measurements and experimental results

The RSS is calculated as

RSS = 10log((Etot · λ

4π)2) (3.9)

3.3 Measurements and experimental results

3.3.1 Outdoor RF propagation verification

The proposed method was verified by comparing the results to that of Munich

scenario [92], which was comprehensively studied by the European COST 231

working group. Fig. 3.7 illustrates the top view of the target region in Munich

city. The size of the region is 2400 m × 3400 m and there are 3 different routes

measured by COST 231 group (e.g. the red color is the first route rout0 named as

METRO200, the black one rout1 is the second route METRO201, and the green one

rout2 is the third rout METRO202).

A new training database is created through similar methods described in Chapter

2. The training images are fetched randomly from Google Maps within the region of

Munich city. Buildings, roads and trees are labeled in order to guarantee the learning

of reliable and transferable knowledge. Fig. 3.8 shows the evaluation procedure: At

the beginning, the same area in which the COST 231 group did measurements is

captured on Google Maps. The images are downloaded and divided into small

sub-images, each of which is of the similar size as the training images.

Afterwards, they are passed through the image classifier to recognize and segment

the objects belong to three classes (building, tree and road). The recognized objects

are then vectorized and the 3D database is automatically generated. Then the

propagation simulator is run to simulate the signal transmission in the target area.

Finally, the signal coverage map is visually presented to the user.

The proposed method is programmed by combining C#, Matlab and C++ code.

It was run on a PC equipped with Intel i5-760 CPU, 2.8 GHz. It took less than

1 hour to construct the 3D database and approximately 15 minutes to finish radio

estimation with a resolution of 1 m. Hence it took around 0.3 s to process 2356

points for all the three routes.

The simulation results are compared with the practical measurements. As shown

in Fig. 3.9 ∼ Fig. 3.11, the results of the 3 different routes correlate well with the

83


Figure 3.7: Three different routes measured by COST231 group.

real measurements. The mean error (ME) and the standard deviation (STD) of the

results are the two metrics used to evaluate the accuracy of the simulation method.

The ME of METRO200 is -0.3 dB and the STD is 5.7 dB. The ME is -0.4 dB and

STD is 5.5 dB in METRO201. The ME and STD of METRO202 are 2.2 dB and

6.7 dB respectively. The performances on the first two routes are similar and are

84


A small region of original image

Classification result

Radio simulation on the constructed 3D scene

3D reconstruction

Figure 3.8: Classification result and radio propagation procedure over Munichscenario.

85


0 200 400 600 800 100070

80

90

100

110

120

130

140

150

160

170

180comparison on METRO200

number of location

path

loss

(dB

)

real measurementsimulated result

Figure 3.9: Simulation result for the first route METRO200.

0 50 100 150 200 250 300 350 40070

80

90

100

110

120

130

140

150

160

170

180comparison on METRO201

number of location

path

loss

(dB

)


Figure 3.10: Simulation result for the second route METRO201.

86


0 200 400 600 800 1000 120070

80

90

100

110

120

130

140

150

160

170

180comparison on route METRO202

number of location

path

loss

(dB

)


Figure 3.11: Simulation result for the third route METRO202.

Table 3.1: Comparison of radio estimation result with other methods.

MethodMETRO200 METRO201 METRO202STD Mean STD Mean STD Mean

This thesis 5.7 -0.3 5.5 -0.4 6.7 2.2

Ericsson 6.7 0.3 7.1 2.3 7.5 1.4

CNET 6.9 -2.1 9.5 -3.6 5.6 -0.2

COST-WI 7.7 10.8 5.9 15.4 7.3 16.3

RAY-TRI [111] 7.1 -2.6 6.2 -0.7 8.3 -1.4

better than the third route which has more measured points and some are located

in the streets with much more buildings.

Table. 3.1 compares the proposed method with other state-of-the-art

technologies. The STD and ME of the proposed method are the smallest on

METRO200 and METRO201, while they are slightly greater than CNET in

METRO202. The experimental results indicate that this novel approach has good

efficiency and accuracy. A quantitative explanation for this improvement can

be directed towards three reasons: the construction accuracy of the environment

database is 82%, the miss-classified pixels are 18%. The resolution of simulation

87


Figure 3.12: Four-layer architecture and physical view of the Cookie node.

is 1 m which is higher than in other methods. Furthermore, the propagation

model is adequate in describing reduced dimensionality environments, and most

of the calculations used in the radio propagation model are based on efficient image

manipulation rather than slow geometrical computation.

3.3.2 Indoor RF propagation verification

In this section, indoor experiments are made to verify the simulation method.

Sensor nodes named �Cookies�[112], developed by CEI-UPM, are used in this

experiment. �Cookies�have modular architecture of four layers as shown in Fig.

3.12. The layers are bonded through vertical connectors which contain all the

signals within the node. As all the layers use the same connectors in the same

position, reusability and interchangeability are much easier for catering different

application requirements. Since the experiments are focused on signal strength

measurement, the communication layer with ZigBee communication protocol is to

be concerned. The communication layer includes a ZigBee module ETRX2 from

Telegesis as shown in Fig. 3.13, which is a low power 2.4 GHz band transceiver based

on the Ember EM250 SoC ZigBee/IEEE802.15.4 solution. The antenna pattern is

near Omni-directional (see Fig. 3.14) with linear polarization.

The transmitter is connected and power supplied by laptop through USB cable,

the transmit power is set to be 3 dBm, the sensitivity of antenna is -97 dBm, the

receiver is powered by two 1.5 V batteries. TX sends packet to RX to request link

88


Figure 3.13: ETRX2 ZigBee communication module on communication layer.

Figure 3.14: Radio pattern of antenna of ETRX2 module.

information. Then RX fetches RSSI value of the received packet and replies to

the transmitter. By placing the receivers at different locations, different RSSIs are

obtained.

The measurement is carried out in the lab of CEI-UPM. Three different scenarios

are measured, as shown in Fig. 3.15. The Scenario A is realized in a room equipped

with several desks and computers. The transmitter is fixed at 1.04 m high above the

ground. The receiver is kept at the same height as the transmitter, and RSSIs are

measured at five locations as indicated in Fig. 3.15(a). The purpose of this scenario

is to verify the propagation model with trivial obstacles, exam the value of path loss

coefficient n and RSS(d0). In scenario B (Fig. 3.15(b)), the position of transmitter is

unchanged, while the locations of receiver are selected further away from transmitter,

and are in different rooms around the lab. This is to verify the attenuation model

89


ZigBee Coordinator ZigBee Sensor Device

(a) Scenario A

(b) Scenario B

(c) Scenario C

112

BS

BS

BS

Figure 3.15: Three scenarios for radio measurements.

of non-trivial obstacles, and obtain empirical penetration parameters. Scenario C

(Fig. 3.15(c)) is realized in the corridor where the communication between TX and

RX are LOS, and because the width is much smaller compared with the length, the

tunneling effect occurs. Not only both transmitter and receiver are maintained at

same height, but also antennas are pointed toward directions with maximum RSSI

to minimize the radiation pattern degradation.

The ray tracing engine is programmed in C++ for Windows operating system.

Fig. 3.16 shows the visible planes (with blue color) and the related information

90


tx

rx

tx

rx

tx

rx

Figure 3.16: Ray tracing demonstrations from different TXs and RXs.

discovered at the polar sweep step. Actually, the floor and ceiling are also visible

planes to TX, but they are not highlighted for providing a friend vision. Those

three figures show all the ray paths from TX (magenta point) to a RX (yellow

point). As can be seen, the number of rays that could arrive at a RX is varied.

Directed, reflected, diffracted rays are computed through full 3D approach without

simplification during ray tracing.

Fig. 3.17 and Fig. 3.18 show the simulated RSS maps for the campaigns.

Both ray tracing algorithm and radio propagation model are developed for 3D

computation. As can be seen, the walls significantly reduce signal strength when

encountered by the signal path. During simulation, some parameters are constant,

the resolution is set to be 0.2 m.

Different groups of simulation results are obtained by varying depth from 0 to

6, which are then compared with the real measurements, as well as the Free-space

model (3.10) and Log-normal model (3.11), where PLref is the path loss at reference

91


Figure 3.17: Simulation result: example 1.

Figure 3.18: Simulation result: example 2.

distance and Xσ is gaussian random value with 0 mean and σ as the standard

deviation which is equal to 3 in this study,

PLfs(dB) = 20 log10 (d) + 20 log10 (f)− 147.55 (3.10)

PLLN (dB) = PLref + 10 nln log10 (d/d0) +Xσ (3.11)

Scenario A is a typical indoor environment without important obstacles between

transmitter and receiver. Fig. 3.19 compares the proposed ray tracing method with

Free-space path loss model, Log-normal model and real measurements. Table. 3.2

shows their mean error ME and standard deviation error STD in dB, compared

with the real measurements. Free-space model is too optimistic in estimating indoor

propagation with the highest ME and STD. While Log-normal model is much

better than the Free-space model. The ray tracing is better than the Log-normal

model. Moreover when depth = 3, the ray tracing method is the most accurate

result. If all the RXs are in the same room of TX, the dominant contributions of

92


1 2 3 4 5-120

-110

-100

-90

-80

-70

-60

-50

-40

-30

Sensor ID

RSS

I (dB

m)

Result Comparison for Scenario A

Direct pathDepth 3Free SpaceLog-normalReal Measurment

Figure 3.19: Results and comparisons of Scenario A.

RSS are direct, reflected and diffracted paths without considering penetration losses.

Some parameters of the proposed radio propagation model are determined through

this case study to calibrate with the real world impacts, thus n = 2, Lobstacle(watts) =

3.6 and α = 0.91. The reference distance loss of the Log-normal model PLref

is calibrated to 60 dB and nln = 3 which will be maintained unchanged for the

remaining two scenarios. One should also notice that the gaussian random variable

Table 3.2: Results comparison: Scenario A.

��ResultMethod Proposed

FS LNdepth = 0 depth = 3

ME (dB) 3.18 1.99 30.87 6.42STD (dB) 3.22 2.00 6.28 6.67

93


1 2 3 4 5-120

-110

-100

-90

-80

-70

-60

-50

-40

-30

Sensor ID

RSS

I (dB

m)

Result Comparison for Scenario B


Figure 3.20: Results and comparisons of Scenario B.

in the Log-normal model may slightly change the performance for the same scenario

at different simulations.

In Scenario B, RXs are in different rooms than TX, all possible paths may make

contributions to RSS. In this scenario, the penetration attenuations have strong

impacts on the RSS, see Fig. 3.20. Table. 3.3 shows the ME and STD for each

method. As expected for the ray tracing method, the results that only consider

direct path are slightly worse than the ones considering multi-path effects. Since

Log-normal model does not consider obstacles between TX and RX, it can not

accurately model the indoor environment when obstacles exist and the ray tracing

method turns out to be the best in this case.

In Scenario C, there is no any obstacle between the TX and RXs. However due

to the narrow width of corridor, reflections and diffractions play more important

94


Table 3.3: Results comparison: Scenario B.



ME (dB) 2.71 2.42 37.29 7.49STD (dB) 2.91 1.65 4.02 4.61

role than the previous two scenarios. As can be seen from Fig. 3.21, the ray tracing

results match quite well with the real measurement even with small fluctuations. The

slope of Free-space model is almost the same as ray tracing, which is reasonable, as a

corridor can be considered as a free space environment in a long distance. However,

the entire indoor environment is definitely not a free space. Table. 3.4 indicates

that the proposed ray-tracing method outperforms the other two algorithms.

1 2 3 4 5 6 7 8 9 10 11 12-120

-110

-100

-90

-80

-70

-60

-50

-40

-30

Sensor ID

RSS

I (dB

m)

Result Comparison for Scenario C


Figure 3.21: Results and comparisons of Scenario C.

95


(a) Bird’s view of the deployment on the toy scenario

(b) Zoom in of the deployment

Figure 3.22: Demonstration of the toy example which has similar deployment asscenario C.

As it can be noticed, the accuracy rises as the value of depth grows. Through

the comparisons of the three scenarios, it is concluded that the proposed ray tracing

method performs more accurately and robustly than other methods. Therefore it is

suitable to be applied for different scenarios in indoor environment.

The ray tracing engine is run in a PC with Intel Core i5-760 CPU of 2.8GHz

frequency. Besides the three deployments in CEI-UPM environment, a virtual WSN

deployment is realized in East Lansing scenario with the topology similar as scenario

C, see Fig. 3.22.

Table 3.4: Results comparison: Scenario C.



ME (dB) 1.19 1.15 30.23 10.65STD (dB) 0.80 0.80 1.41 1.41

96


direct 1 2 3 4 5 60

5

10

15

20

Reflection depth

Tim

e (s

)

Scenario AScenario BScenario CEast Lansing

Figure 3.23: Average time consumption of polar sweeping, by varying the depth ofreflection rays.

The average time consumptions of polar sweeping and ray tracing, by varying

the reflection depth for each tested scenario, are shown in Fig. 3.23 and Fig. 3.24

respectively. Both figures indicate that the computation time increases as depth

increases. The time consumptions of polar sweeping for scenarios A∼C in CEI-UPM

are quite close to each other, while it took 2.8 times longer in average to finish the

polar sweeping in East Lansing environment. The time cost of ray tracing of

East Lansing goes much higher than that of Scenario C when depth exceeds 3, as

much more new reflection planes are involved in the former environment. Fig. 3.25

shows that the polar sweeping without kd-tree traversing takes 3.35 times longer

than the opposite choice, therefore the kd-tree traversing can significantly improve

the efficiency of the ray tracing engine.

The engine takes 37 ms in average to construct a kd-tree for CEI environment

and 78 ms by doing so for the East Lansing environment. Note that the kd-tree

construction is realized based on splitting properly all the primitives in the target

region, the time spent on kd-tree construction is dependent on the number of planes

in the environment. While the time spent on tracing rays and polar sweeping

97


direct 1 2 3 4 5 60

5

10

15

20

25

30

35

40

Reflection depth

Tim

e (m

s)

Scenario AScenario BScenario CEast Lansing

Figure 3.24: Average time consumption of ray tracing, by varying the depth ofreflection rays.

depends on the value of depth and the complexity of surrounding environment of

the transmitter. The more reflections and planes exist around a TX, the longer the

simulation time will be.

Table 3.5: Attenuation parameters of major objects indoors.

Materials Brick GlassMetallic board

of PC

Human Body

Antenuation

Parameter (dB)8 2 5 2∼10

During the experiment, the attenuation parameters of different obstacles are also

measured along the propagation path, as indicated in Table 3.5. The main obstacles

are walls, glass of windows, metallic boards and human bodies. The attenuation

caused by stilled human body even varies widely. Human movements are dynamic

and there are some researchers focusing on the interference of human body on the

radio propagation. The significance of the effect of human body towards signal

strength depends on how high the sensor nodes are installed, and whether human

98

3.4. Conclusion

direct 1 2 3 4 5 60

5

10

15

20

Reflection depth

Tim

e (s

)

polar sweep with kd-treepolar sweep without kd-tree

Figure 3.25: Average time consumption of polar sweeping with and without kd-treetraversing.

movement is important for sensor applications, such as in localization systems or

health care applications.

3.4 Conclusion

The developed ray tracing engine propagation simulation method contributes on

efficiency and accuracy to the radio estimation. By using image processing concepts

including the kd-tree space division algorithm and modified polar sweep algorithm

the rays are traced efficiently without detecting all the primitives in the scenario.

The ray-tracing algorithm ensures accurate and practical results. The radio

propagation model emphasizes not only the materials of obstacles but also their

locations. Hence, the performance of simulation is robust and accurate compared

with conventional propagation models. The experimental results imply that this

methodology is suitable for both outdoor urban scenes and indoor environments,

moreover it can be applied to GSM communication and ZigBee protocol by varying

99


frequency parameter in the model. Moreover, the ray tracing engine is featured with

generality that can be used for any environment database with the similar description

format on primitives. The experimental data are available in this chapter, which can

be used for comparisons by other researchers. The time consumptions are recorded

for all the steps of ray tracing, the measurements on RSSs for different deployments

and the attenuation parameters of different materials are also available in this work.

Beyond the RF propagation estimation, the sensing signal of sensor nodes, which

are sensitive to the obstacles, can also be benefit from the ray-tracing engine.

The details for modeling the sensing coverage is introduced in the next chapter

together with the planning algorithm. The indoor furniture also have influences on

RF propagation and sensing signal, although it is not introduced in this chapter,

with the ray tracing method, they can be detected. Therefore once provided with

their models together with the environment database, the calculated results will

automatically consider their impacts without any simplification.

100

Chapter 4

Planning the WSN

One of the major challenges in designing wireless sensor networks is the support

of various application requirements while coping with the computation, energy,

communication, sensing and cost constraints. Careful node placement can be a

very effective optimization means for achieving the desired design goals. However,

optimal node placement is a very challenging problem that has been proven to

be NP-Hard for most of the formulations of sensor deployment [59, 60, 61], and

the modelling of important metrics turns out to have significant impacts on the

deployment decision.

In this chapter, design parameters including the network topology, heterogeneity

of nodes and the modelling of important metrics are discussed. The current

researches on optimized node placement have been reported in Chapter 1. According

to the reviews, those algorithms and heuristics somehow have serious limitations.

From the aspect of metric modelling, only a few of them tackle the 3D deployment

issue and they are developed only for 3D indoor applications [42, 51, 52]. Even fewer

works model the sensing coverage and radio propagation by considering the realistic

scenario where obstacles exist. None of the aforementioned algorithms modeled

the network longevity properly and practically, they often employ unilateral and

unrealistic formulations. Moreover, the optimization targets are one-sided. Without

comprehensive evaluation on the important metrics, the performance of WSNs can

not be entirely optimized and reliable.

An efficient WSN planning algorithm is proposed in this work to tackle the above

mentioned challenges and efficiently assist designers on deploying reliable WSNs.

This algorithm contributes on the following aspects:

• Comprehensive metrics are considered. This work considers

connectivity, sensing coverage, cost, lifetime, packet delay and packet loss rate,

Chapter 4. Planning the WSN

which to our best knowledge, is the most comprehensive evaluation scheme for

analyzing the performance of WSN.

• Practical metrics modelling by integrating network simulator. The

connectivity and sensing coverage are modelled in assistance of 3D ray-tracing

method which is sensitive to the existence of obstacles; hardware cost refers to

the number of devices as well as their types. Routes of network are constructed

by using AODV protocol based on the computed connectivity information.

Network longevity, packet delay and packet drop rate are obtained through

triggering events in WSNet simulator according to a user defined sensing task

and the provided topology. It is the first time that network simulator is

involved in a planning algorithm to tackle the difficulty on modelling those

vital metrics and provide practical evaluations.

• Efficient and multi-objective optimization. A multi-objective

optimization algorithm is developed for WSN to optimize the cost, coverage,

lifetime, packet delay and packet drop rate. The individual length is

changeable so that the cost can be optimized, meanwhile crossovers and

mutations are designed to eliminate invalid modifications to improve the

computation efficiency. NSGA-II ranking method, which is proved with high

efficiency, is employed by this work.

4.1 Topology

One important property of a sensor network is its diameter, that is, the maximum

number of hops between any two nodes in the network. In its simplest form, a sensor

network forms a single-hop network, with every sensor node being able to directly

communicate with every other node. An infrastructure-based network with a single

base station forms a star network with a diameter of two. A multi-hop network may

form an arbitrary graph, but often an overlay network with a simpler structure is

constructed such as a tree or a set of connected stars. The topology affects many

network characteristics such as latency, robustness, and capacity. The complexity

of data routing and processing also depends on the topology. Fig. 4.1 shows four

different topologies that have been applied to WSN applications.

102

4.1. Topology

SN

BS

BS

BS

CH

CH

CH

BS

CH

CH

SN

SNSN

(a) Star Network (b) Tree Network

(c) Mesh Network (d) Cluster Network

Figure 4.1: Different topologies of WSN.

4.1.1 Star network

In star networks, nodes are connected to a centralized communications hub. Each

node cannot communicate directly with one another; all communications must be

routed through the centralized hub. Each node is then a�client�while the central

hub is the �server�itself or a gateway node that is in direct communication with

the base station. An example of a star network is shown in Fig. 4.1(a). The failure of

a link does not affect the entire network and leaves the rest of structure unchanged.

However, it is not fault tolerant as there is no alternative path from unconnected

node to the base station once the link is failed or obstructed. Due to the constraint

on radio communication ability, sensor nodes in star topology can not be placed far

away from the sink, thus its applications are limited to small networks. As there is

no delay due to buffering at routers along the path, the data latency in star topology

103


is quite low, nevertheless there is likely to be more loss due to collision when network

density increases.

4.1.2 Tree network

A natural and logical extension of the star topology is the tree structure where sink

node is the root and a collection of star networks are arranged at different levels in

hierarchy, see Fig. 4.1 (b). All the communications between children (lower level)

must be routed through their parent (upper level). The parent node has higher traffic

load thus the energy is depleted faster than its children. Once die out of battery,

the children attached to it will be affected and become isolated. Accordingly, as the

level of node becomes higher, more nodes will be influenced once the critical link is

broken.

4.1.3 Mesh network

Mesh Networks are multi-hop local area networks in which each sensor node not

only sends and receives its own message but also functions as a router to relay

messages for its neighbors through the network. Mesh topology facilitates multiple

communication paths from the sensor nodes to the base station. A special case of

mesh is the grid topology where each grid point represents a sensor node, and the

links are the edges of the grid. An example of mesh topology is shown in Fig. 4.1

(c).

Multi-hop routing methods effectively overcome shadowing and path loss effects,

thus mesh WSNs are self-configuring networks that dynamically optimize routes

through the network based on the best link quality between neighbor nodes. If the

node density is high enough, multi-hop routing uses a large amount of nodes to

balance the relay traffic all over the WSN. However data latency increases as the

number of hops increases, WSN designers should carefully plan the transmission and

duty-cycle scheduling to overcome the collision and interference. Generally speaking,

the amount of data, which are relayed for each node, increases as a node getting

closer to the base station, therefore some redundant nodes are necessary to backup

in the bottleneck area of WSN. Many applications employ mesh topology, such as

104

4.2. Heterogeneity

the one for monitoring active volcano [17] and the GreenOrbs deployment [113] for

ecological surveillance in the forest.

4.1.4 Cluster network

In clustered hierarchical topology, all nodes in a WSN are joined at the lowest level.

The sensors use their local neighborhood information to form a set of clusters and

elect a cluster head (CH) for each cluster. The CH election process can be based on

various parameters such as available energy resources, proximity to the base station,

and number of neighbors. The CH in the lowest level are arranged into clusters in

a higher level. The process is repeated for each level in the hierarchy. The number

of hierarchical levels depends on several criteria including coverage requirement,

deployment region, node density, and transceiver and sensing range. The clustered

hierarchical architecture maintains a tree routed at the sink node, with a hierarchy

of CH as the internal nodes and sensor nodes as leaf nodes of the tree. Nevertheless,

different from the tree topology, it still maintains the multi-hop mesh within the

same cluster.

Compared with mesh mode, data aggregation is more efficient in cluster topology

as only the CHs are in charge of aggregating and forwarding data to the sink, hence

it can be extremely effective in one-to-many, many-to-one, or one-to-all (broadcast)

communications.

On the other hand, all data directed to the sink results in the CH near the sink

come to have high relay traffic. As a result, CH around sink uses much more energy

compared with other CHs, and this is one of the main problems that shorten the

network life time in clustered WSN. Due to its complexity in CH election, each

node should be aware of status of its neighbors and to our best knowledge, cluster

topology has not been practically employed by any WSN application.

4.2 Heterogeneity

When considering the heterogeneity, WSN can be classified into two categories:

homogeneous WSN and heterogeneous WSN. Early sensor network visions

anticipated that sensor networks would typically consist of homogeneous devices

that were mostly identical from a hardware and software point of view. Some

105


projects, such as Amorphous Computing [114], even assumed that sensor nodes

were indistinguishable, that is, they did not even possess unique addresses or IDs

within their hardware. This view was based on the observation that otherwise it

would not be feasible to cheaply produce vast quantities of sensor nodes.

However, in many prototypical systems available today, sensor networks consist

of a variety of different devices. Nodes may differ in the type and number of attached

sensors [32, 44]; some computationally more powerful nodes may collect, process,

and route sensory data from many more limited sensing nodes [45]; some sensor

nodes may be equipped with special hardware such as a GPS receiver to act as

beacons for other nodes to infer their locations such as the projects in [115, 116],

which use GPS and satellite to locate the animals; some nodes may act as gateways

to long-range data communication networks (e.g. 3G gateways are used in [117],

satellite networks, or the Internet [118]). The degree of heterogeneity in a sensor

network is an important factor since it affects the complexity of the software executed

on the sensor nodes and also the management of the whole system.

4.3 Introduction and modelling of important metrics

4.3.1 Preliminaries and assumptions

We summarize the important symbols that are used in this section in Table 4.1. Our

Table 4.1: Important symbols

Symbol Meaning

Si sensor node with ID = i

Ni node device with ID = i and S ∈ N

As sensing/monitoring area

ΦSi a set of points that are the covered points of Si

C coverage

cost cost

L lifetime

Pl packet latency

Pd packet drop rate

DC,cost,L,Pl,Pddesirability formulation of C, cost, L, Pl, Pd

106

4.3. Introduction and modelling of important metrics

iSiSr

),(ISi rSsphere

)( imSO

iSm ��

Figure 4.2: The searching of covered point

method employs a ray-tracing method described in our previous work [109] as well

as in Chapter 3 to emulate the propagation paths of both radio and sensing signals.

In the real world propagation, multi-path phenomena occurs on the radio signal

transmission: when a radio signal encounters obstacles, reflections and diffractions

happen. Thus received signal strength at the receiver is computed by accumulating

the arrived waves from all directions. Unlike the radio propagation, the sensing

signal is usually only considered by using direct path, which halts at the intersected

points with the surfaces of obstacles. Thus a covered point of a sensor node Si is

defined as following:

Definition 1 Covered point: A point m is said to be covered by Si, if and only

if it is within the sensing range of Si and is not obstructed by any obstacle. ΦSi

represents a set of all the points that are the covered points of Si.

ΦSi = {m|m ∈ sphere(Si, rSi)∧!O(−−→mSi)}

As indicated in Fig. 4.2, sphere(Si, rSi) is the sphere with radius rSi centered at

node Si and indicates the ideal sensing area of Si. O(−−→mSi) indicates whether the

sensing path from m to Si is obstructed.

107


Routing scheme: There have been many algorithms proposed for routing data

in sensor networks, which consider the characteristics of sensor nodes together with

the application requirements. Nowadays, an overwhelming number of commercial

sensor devices support the distance vector based routing protocol such as the Ad

hoc On-Demand Distance Vector (AODV) routing [119] and link-state based routing

protocol Dynamic Source Routing (DSR) [120]. However, the routing results of

minimum-weight/minimum-hop based routing protocols are very similar and can be

computed based on the shortest-path searching by Dijkstra’s algorithm, in which

data are collected and forwarded to BS via the path with the best distance metric.

In this work, the distance metric is modelled by the qualities of established links

according to the ray-tracing results, which not only considers the distance between

transmitter and receiver but also indicates the impact of surrounding environment.

According to the comparison and traversing procedure over the WSN, the pseudo

code of Dijkstra algorithm is shown in Listing. 4.1.

Listing 4.1: Pseudo code of Dijkstra algorithm

S={all nodes except the base station V};

For each node u{/*search the path from u to V*/

int temp = maxint;

int u = v;

For each node j in S {

/*search a un-traversed node with shortest path to V*/

if ((!S[j])&&(dist[j]<temp)){

u = j;

temp = dist[j];

}

}

S[u] = true; /*Remove u from the list*/

For each node j {

if ((!S[j])&& j is directely connected wih u){

int newdist = dist[u][v] + max(c[u][j], c[j][u]);

/*if the path through u is smaller than the older

distance from j to v

the shottest route is updated for node j*/

if (newdist<dist[j][v]) {

dist[j][v] = newdist;

108


next[j] = u; /*the next hop of j is u */

}

}

}

}

4.3.2 The cost of WSN

One of the design goal, from economic point of view, is to reduce the cost while

fulfilling requirements of an application. Many companies and research organizations

arise in the recent decade to design and manufacture sensor nodes, which provide

various options on the budget. Besides, the installation of sensor nodes requires

extra human efforts. For instance, the cost of placing sensors on the ceiling and

walls is different from placing them within a human-active space; even attaching

sensor nodes to different heights can vary the costs. As a result, the Cost of WSN

is categorized into hardware cost and deployment cost in this work.

4.3.2.1 Hardware cost

MICA and MICA2 were once the most successful families of Berkeley motes. The

MICA2 platform, whose layout is shown in Fig. 4.3, is equipped with an Atmel

ATmega128L and has a CC1000 transceiver. Berkeley motes up to the MICA2

generation cannot interface with other wireless-enabled devices [121]. However,

the newer generations MICAz and Telos support IEEE 802.15.4, which is part of

the 802.15 Wireless Personal Area Network (WPAN) standard being developed by

IEEE. At this point, these devices represent a very good solution for generic sensing

nodes, even though their unit cost is still relatively high (about $100 ∼ $200).

Various platforms have been developed for the use of Berkeley motes in mobile

sensor networks to enable investigations into controlled mobility, which facilitates

deployment and network repair and provides possibilities for the implementation

of energy-harvesting. UCLA�s RoboMote [122], Notre Dame�s MicaBot [123]

and UC Berkeley�s CotsBots [124] are examples of efforts in this direction.

UCLA�s Medusa MK-2 sensor nodes [125], developed for the Smart Kindergarten

project, expand Berkeley motes with a second microcontroller. An on-board power

109


management and tracking unit monitors power consumption within the different

subsystems and selectively powers down unused parts of the node.

Intel has designed its own iMote [126] (see Fig. 4.3) to implement various

improvements over available mote designs, such as increased CPU processing power,

increased main memory size for on-board computing and improved radio reliability.

In iMote, a powerful ARM7TDMI core is complemented by a large main memory

and non-volatile storage area; on the radio side, Bluetooth has been chosen, thus

the price of an iMote is about $299 which is very high.

MICA2 TelosB iMote2

Figure 4.3: Products by Crossbow.

The BTnode rev3 hardware designed in 2007 by ETHZ [121] (see Fig. 4.4),

is based on an Atmel ATmega128L microcontroller, a Bluetooth module and a

low-power radio which is the same as that used on Berkeley MICA2 Motes. BTnode

costs around $200 which is slightly higher than MICA2 due to an extra cost on

Bluetooth module.

Figure 4.4: BTnode rev3.

110


Waspmote [127] in Fig. 4.5 is a sensor device developers oriented product. It

works with different communication protocols (ZigBee, Bluetooth and GPRS) and

frequencies (2.4GHz, 868MHz, 900MHz) and creates links with distance up to 12

km. Waspmote is compatible with more than 50 sensors. The flexible property

makes Waspmote suitable for different types of applications and the price varies

according to the configuration. For example, sensors with ZigBee module cost the

least compared with that attached with Wi-Fi and Bluetooth module, while the one

with 3G+GPS costs around $300 and is the most expensive configuration.

WiSMote [128] in Fig. 4.5 appears since 2011. It is a sensor/actuator

module well adapted to WSN applications. The wireless link operates over the

2.4 GHz ISM band. With its wide range of embedded sensors and its variety of

extension connectors, it is able to monitor any kind of physical measurements in

fields like environment, healthcare, domotics, smart building, logistics or industrial

applications. WiSMote embeds an small footprint operating system (Contiki) plus

an IEEE 802.15.4 protocol stack compatbile with Zigbee and 6LoWPAN (IPv6).

The availability of inexpensive hardware such as CMOS cameras and

microphones has fostered the development of Wireless Multimedia Sensor Networks

(WMSNs), i.e., networks of wirelessly interconnected devices that are able to

ubiquitously retrieve multimedia content such as video and audio streams, still

images, and scalar sensor data from the environment. One of the recent multimedia

wireless sensor is the SEED-EYE [129] in Fig. 4.5, which provides an advanced

board for implementing low-cost (about $150 per unit)WMSN. It hosts a powerful

Microchip PIC32, and has a full set of communication interfaces such as Ethernet,

IEEE802.15.4 / Zigbee, and USB. Moreover, it is integrated with a CMOS Camera,

making it an ideal board for implementing next generation imaging WSNs.

Beyond the on-shelf sensor products, there are enormous research organizations

dedicating to develop and prototype their own sensor platforms. Typically, those

prototypes are based on add-on/modular hardware design. Although the designs

are not highly optimized in terms power consumption or size and price, one can still

foresee the bright future once they are launched in the market. The Tyndall’s mote

family [130] developing a range of ISM band wireless sensing systems for deployment

in the environment focusing on the key areas of the environment and fitness and

health, the structure is shown in Fig. 4.6; �Cookies�[112] developed by the

111


Waspmote Wismote SEED-EYE

Figure 4.5: Waspmote(left),Wismote(middle), SEED-EYE(right).

researchers at Centro de Electronica Industrial of Universidad Politecnica de Madrid

(CEI-UPM), have a modular architecture of four layers, as depicted in Fig. 3.12.

Each layer fulfills a specific functionality in the node, and the layers are changeable

for different applications. Moreover, it is possible to have a heterogeneous network

with nodes composed of different layers. In the current motes, 2.4 GHz ZigBee

communication protocol is used and AODV protocol is embedded.

Tyndall Mote

Figure 4.6: Tyndall mote.

During the investigation of different sensor node manufacturers, we find out

that some of them also produce other types of node to cater the heterogeneity and

various applications of WSN. WiSGate of Arago Systems and Meshlium produced by

Libelium act as a gateway between a wireless sensors/actuators network and other

type of network. In [131], the authors analyzed hybrid sensor networks consisting

of transceiving (cluster-heads) and transmit-only sensors. By using the developed

112


mathematical model of physical and MAC layer, they demonstrated how much the

dollar-cost and the power consumption of a sensor network can be decreased while

maintaining the same network coverage.

Table 4.2 summarize the hardware features of above mentioned sensor

productions and compare their prices. Due to the heterogeneity property and various

topologies of WSNs, motes with different functions should be considered to obtain

a optimum cost solution rather than simply using a uniform type of all-function

sensor mote for the whole network. Based on the above survey, motes are classified

into three types in this work:

• Sensor Node (SN): equipped with sensors to monitor the surrounding

environment. In this work, each sensor node is static and has wireless

communication and routing ability.

• Relay Node (RN): has the ability of communication and routing. RN is usually

needed to fill radio communication hole or to balance traffic load.

• Base Station (BS): is in charge of aggregating data and is directly connected

with the central server. A WSN has one BS and its location is predetermined

by users.

Since the number and location of BS is assumed to be fixed in this work, the cost

of BS is also fixed and is not included in the model of hardware cost. SN contains

extra sensor module besides the communication and routing module, therefore a SN

costs more than a RN. The hardware cost model costhw is expressed by accumulating

the price of each node deployed in the area A:

costhw =

M∑i=1

(P (Ni.type)) (4.1)

where P (Ni.type) indicates the relative price that is dependent on the type of node

Ni, for instance P (RN) = 1, P (SN) = 3.

The available budget is considered as the maximum cost costmaxhw. The

normalization of costhw provides a desirability component of the design goal to

minimize the hardware cost, and Dcost is used in this work to represent the

113


Tab

le4.2

:Features

ofvario

usplatfo

rm.

Platfo

rmYear

CPU

Communicatio

nExtern

al

Memory

Power

Supply

Pric

e

MIC

A2

2002ATMega128L

CC1000

512kB

Flash

2×AA

$150

Telo

s2004

MSP430F149

CC2420

512kB

Flash

2×AA

$110

Tmote

Sky

2005TIMSP430

CC2420

1M

Flash

2×AA

$99SunSpot

2006

AT91RM920T

CC2420

4M

Flash

Lith

ium

ion

$750

Imote2

2007Intel

PXA271

CC2420

32M

Flash

2×AA

$299

BTnoderev

32007

ATMega128L

Blueto

oth

CC1000

128KB

Flash

2×AA

$215

Wasp

mote

2009ATmega1281

8radio

modules

128KB

Flash

2×AA

$130˜$30

0W

iSMote

2011

MSP4305x

CC2520

2M

Flash

2×AA

N/A

SEED-E

YE

2012PIC

32MX795F512L

MRF24J40MB

N/A

N/A

$200

114


desirability on hardware cost.

Dcost =costmaxhw

− costhwcostmaxhw

(4.2)

4.3.2.2 Deployment cost

During the deployment of nodes and equipments, human efforts and extra tools are

needed in order to mount the nodes in the decided locations. The investigation on

the lessons learnt by other researchers indicates that mounting problem occurs in all

the works. The authors of [132] deployed WSN for precision agriculture, and they

observed that multi-path fading which was exacerbated by the movement of leaves

of the maize plants played a very crucial role on RSSI. From 2004-2005, Langendoen

et al. [133] tried to deploy more than 100 sensor nodes to monitor the potato crops

for precision agriculture. To avoid obstruction of when the potato crop is flowering

and leaves cover the (ground-based) antennas, the nodes were installed on poles

at a height of 75 cm. Besides, they included a safety margin to ensure that the

nodes could not be hit by farming equipment attached to a tractor. The authors

also learned very interesting lessons in [134], they developed a WSN to monitor the

indoor environment. However, they found in two occasions that the sensor nodes

are taken from where they were. As a result, they decided to place the sensor nodes

not for the best coverage but for the best security. Hence those sensor nodes are

eventually either hidden from field of vision or placed high up on the wall.

There might exist non-deployable area where nodes are not allowed to be placed,

those area should be pre-defined and the corresponding deployment cost can be

assigned with the maximum value. In this work, deployment cost costd is modeled

as boolean value and the forbidden area is assigned by users through GUI interface

as shown in Fig. 4.7, costd = 1 if the area is not accessible, otherwise costd = 0.

4.3.3 Coverage

Sensing coverage is one of the key issues that should be considered when deploying a

WSN, as it corresponds to the quality of service that can be provided by a WSN. The

coverage concept can be defined and categorized based on the node density level.

If only parts of the area are covered by the sensor nodes, the coverage is sparse;

if the area is completely (or almost completely) covered by sensors, the coverage

115


Forbidden Area

h=2 mh=1 m

Width=3 m

cost=0

cost=0

cost=1cost=1

Figure 4.7: Deployment cost configuration in vertical view.

is dense; otherwise, if the same detected location is covered by multiple sensors,

the coverage is called redundant. The density of coverage is normally determined

by user requirements which may vary across different applications. An adequate

coverage is a key to robust WSN application, and it may also be exploited to extend

the network lifetime by switching redundant nodes to sleep modes to reduce power

consumption. In this work, K-coverage problem is investigated and is concerned as

the potential problem to be tackled by planning algorithm.

Definition 2 Target K-coverage Requirement: At any given moment, any

target point m ∈ As is the covered point of at least k different SNs (k = 1 · · ·M).

Therefore the desirability of K-coverage requirement C is expressed as:

DC =|∑M

i=1{m|m∈As∧m∈ΦSi}|

|{m|m∈As}|×k(4.3)

4.3.4 Connectivity

Wireless sensor networks are represented by a graph G = (V,E) where V is the set

of nodes and E ⊆ V 2 is the set of edges: (Nu, Nv) ∈ E means that Nu and Nv are

neighbors. The neighborhood set N(u) of Nu is expressed by

N(u) = {Nv|(Nu, Nv) ∈ E ∨ (Nv, Nu) ∈ E}

116


wireless links are determined according to received signal strength calculated by

accurate ray-tracing method, thus edges are defined as :

E = {(Nu, Nv) ∈ V 2|u �= v ∧RSS(uv) ≥ RXs}

Where RSS(uv) is the received signal strength from Nu to Nv, and RXs is the

sensitivity of antenna at the receiver.

As can be seen from the edge definition, communication links can be established

if the RSS is above the sensitivity of antenna. A WSN is said to be connected, if

any two nodes belong to a WSN are linked together by edge(s) via single hop or

multiple hops. Connectivity is intermittent if the network is occasionally partitioned.

If nodes are isolated most of the time and enter the communication range of other

nodes occasionally, the communication is said sporadic ([135]). Note that despite

the existence of partitions, messages may be transported across partitions by mobile

nodes, which is not the case in this work where only static network is considered.

Connectivity mainly influences the design of communication protocols and methods

of data gathering. Generally speaking, the this concept can be categorized into two

directions:

Definition 3 Connected K-Coverage Problem: Given a sensor network

consisting of n sensors and an interest region, the network should satisfy the

following two conditions at any moment:

1. Satisfy the conditions of the K-Coverage requirement

2. The communication graph G is connected

Definition 4 K-connected Problem: A graph G is said to be k-connected if

for each pair of vertices there exist at least k mutually independent paths of edges

connecting them. In other words, the graph G is still connected even after removal

of any k − 1 vertices from G.

In [136], three localized K-CDS construction protocols were proposed. The

first one is a probabilistic approach which is based on K-Gossip. The second is

a deterministic approach which is an extension from the K-coverage condition. The

last one is Color-Based K-CDS Construction. The authors in [137] proposed two

117


algorithms: one centralized algorithm CGA and one distributed algorithm DDA,

to tackle the limits of k = m in [136]. Although the protocols and theorems

have been focused for redundant connection for a robust WSN, the practical

heuristic to construct a k-connected WSN is rarely studied. The work in [138]

proposes deployment patterns to achieve full coverage and three-connectivity, and

full coverage and five-connectivity under different ratios of sensor communication

range (denoted by Rc) over sensing range (denoted by Rs) for WSNs. The authors

also discover that there exists a hexagon-based universally elemental pattern which

can generate all known optimal patterns. However, as stated in the paper, when

considering the non-disc sensing model or geographical constraints or heterogeneity

of sensor nodes, the proposed deployment pattern is not optimal.

In this work, connectivity of WSN is constructed to tackle the Connected

K-Coverage Problem with K-connected network topology by using the

previous mentioned sensing model, practical multi-path radio propagation model

on heterogeneous WSNs.

4.3.5 Lifetime, Packet latency and Packet drop rate

4.3.5.1 Lifetime

As most of the sensor nodes are powered by batteries, they will exhaust energy

after a certain time once deployed in the environment. Therefore the WSN will be

disconnected and no longer satisfy the sensing requirement. The lifetime of sensor

network is a very important metric and WSN designers have done many efforts

to prolong it. The methods include developing proper MAC periods, optimizing

the topology to reduce bottleneck nodes, developing back up plans and routing

algorithms. The network lifetime in our work is defined as the time that the first

node exhausts its energy. The desirability of this metric is expressed by a ratio

between the actual lifetime (L) and the maximum expected lifetime (Lmax) by the

WSN designer. The expression of DL is:

DL =L

Lmax(4.4)

118


4.3.5.2 Packet latency

Packet latency is defined as a average end-to-end delay from the source to the

destination (BS). There are many factors that affect packet latency and the most

important factors are: the usage of channel, the hops between source and destination

and the scheduling of nodes along the routing path. The desirability of packet

latency is expressed as:

DPl= 1−

∑Mp

i Pl(i)

Mp(4.5)

where Mp is the total number of data packets generated by all the sensors and Pl(i)

is the latency of packet i.

4.3.5.3 Packet drop rate

Packet drop can be caused by signal degradation over the network medium due to

multi-path fading, channel congestion, corrupted packets rejected in-transit, faulty

networking hardware, faulty network drivers or invalid routes. The packet drop rate

is a ratio between the number of dropped packet (Pd) and the number of generated

data packets, and the desirability over this metric is:

DPd= 1− Pd

Mp(4.6)

4.3.5.4 Proposed strategy by using WSNet simulator to model L, Pl and

Pd

Because L, Pl and Pd can be affected by network topology, real-time communication

and packet load, it is difficult to precisely model them through simple formulas. That

is the reason why protocol designers usually estimate such performance through

network simulator, where WSN can be simulated with a determined or random

topology with nodes being scheduled and data load being assigned based on user

specifications.

Inspired by those works, we integrate WSNet1 simulator in this work to observe

the three complex metrics through practical network simulation. WSNet simulator

allows researchers to analyze WSN performance based on network configuration

1http://wsnet.gforge.inria.fr/

119


which can be read from the files that indicate topology, routing and connectivity

information. Users can either use the embedded provided layer modules or extend

the operation by developing customized modules that will be used in the real

application. The network protocol can be the IEEE 802.15.4 standard for physical

layer and MAC layer, or other desired protocols. The energy model should be

programmed previously to imply how energy is consumed for transmitting/receiving

data packets and for different status such as wake, sleep and idle. The key functions

of energy model are in Listing 4.2.

Listing 4.2: Key functions of energy models in WSNet simulator

/* Battery model*/

/* The editable energy comsumption function for

*transmitting and receiving packets, as well as the idle state

*/

void consume_tx(id,duration);

void consume_rx(id,duration);

void consume_idle(id,duration);

/*802.15.4 MAC layer key functions*/

int check_channel_busy(id); /*check the status of channel for a

node*/

int state_machine(id); /* the state assignment for a node*/

As a result the strategy of the modelling method can be described by the

flowchart in Fig. 4.8. When a candidate topology is generated, the node location,

routing and connectivity files are created in a shared folder between the planning

algorithm engine and WSNet simulator. A ’xml’ script is created to configure the

property of WSN for the simulation in WSNet, including the network size, region

scale, network protocols, energy consumption models and directory of the generated

output files. After WSNet finishing the simulation, it returns the values of network

lifetime, packet latency and packet drop rate to the planning algorithm engine, for

analyzing desirability values by the aforementioned formulas (4.4, 4.5 and 4.6).

120

4.4. The proposed multi-objective optimization methodology

Log files: connection, routing table, location

User RequirementsXml configuration file

Lifetime

Packet latency

Packet drop rate

Figure 4.8: Modelling of L, Pl and Pd by using WSNet simulator.

4.4 The proposed multi-objective optimization

methodology

Once provided application requirements and the deployment environment model,

the problem of planning a WSN is formulated as: Determine the topology of

the network to maximize the 5 desirability values calculated by (4.2 ∼ 4.6).

This is a multi-objective optimization problem, which is proven to be NP-hard.

Multi-objective optimization genetic algorithms are proposed to effectively and

efficiently solve the NP-hard problems, among which the Nondominated Sorting

Genetic Algorithm II (NSGA-II) by [139] is an ideal approach that features with

elitism selection, high computation efficiency O(MN2) (where M is the number of

objectives and N is the population size) and does not need to specify the sharing

parameter. As a result, the multi-objective optimization method in the present

work is based on NSGA-II. It concerns the constraints and rules on formulating the

genes as well as the mutations and crossovers, to cater to the features of WSN, and

therefore the results are achieved efficiently and effectively.

Fig. 4.9 shows an overview of the proposed planning method. Network

deployment is generated based on the deployment constraints and user requirements

on the detected regions and forbidden regions, the value of cost is obtained at this

step. There are three ways to create the deployment:

1. User can determine the locations and properties of nodes manually via the

GUI interface;

121


Network deployment generation

Topology generation

Radio and sensing coverage analysis

User Requirements

Detected region

Deployment constraints

Log filesXml file

Locations

Detection load

Cost

CoverageMulti-Objective

Evaluation Function

Feedback Loop

lP

dP

L

Figure 4.9: The strategy of the proposed planning method.

2. Nodes can be randomly generated based on the constraints;

3. Crossover and mutation modify the node properties during the evolutionary

strategy.

Thereafter, radio propagation and sensing signals are computed for each node by

using the ray-tracing method so that the connectivity and sensing coverage are

obtained, and the topology of the WSN is constructed according to the routing

protocol pre-defined by WSN designer. As discussed in section 4.3, the ’xml’ file

and log files are generated to trigger WSNet simulator so that lifetime, packet latency

and packet drop rate are analysed after WSNet finishing simulation. With all those

metrics provided, the multi-objective evaluation method computes the objectives

(desirability) and selects those candidates with best performance base on NSGA-II,

and the selected population are fed back to the network deployment generation

function to create new populations, thus the objectives are gradually progressed. At

the end, the algorithm can provide multiple elitist solutions to WSN designers.

In genetic algorithm, a candidate solution can also be called individual, creature,

or phenotype. Each candidate solution has a set of properties (its chromosomes

122


or genotypes) which can be mutated and altered. Traditionally, individuals are

represented in a vector of binary value, but other representations are also possible.

The evolution usually starts from a population of randomly generated individuals

and in each generation, the fitness value of every individual in the population is

evaluated. The individuals with better performance are selected from the current

population, and each individual’s genome is modified (recombined and possibly

randomly mutated) to form a new generation to be evaluated in the next iteration.

Commonly, the algorithm terminates when either a maximum number of generations

has been produced, or a satisfactory fitness level has been reached.

In this work, an individual is expressed as DV represented by (4.7), where Ni is

the node with ID = i and it is a ”chromosome” of DV . M is the number of nodes,

which also indicates the size of WSN. Each ”chromosome” has properties including

3D location of the node, type of node, transmission power Ptx, radio sensitivity RXs

and sensing range Rsense. We assume that M , Ni.location, Ni.type can be modified,

as a result this method is based on changeable length which will bring difficulties to

crossover and mutation. Besides the locations and type must be modified according

to certain rules to construct a valid WSN. The details of the strategy are introduced

and problems are tackled later in this section.

DV = [N1, N2, · · ·NM ] , Ni =

⎧⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎩

location : x, y, z

type : BS, SN,RN

Ptx, RXs

Rsense

⎫⎪⎪⎪⎪⎪⎬⎪⎪⎪⎪⎪⎭

(4.7)

The algorithm defines two types of population: Parent and Children, for storing

the parents and children respectively. The format of population is defined by Np

individuals attached with their corresponding desirability values:

population =

⎡⎢⎢⎢⎢⎢⎣

DV1, D1cost,c,L,Pl,Pd

DV2, D2cost,c,L,Pl,Pd

· · ·DVNp , D

Np

cost,c,L,Pl,Pd

⎤⎥⎥⎥⎥⎥⎦ (4.8)

123


4.4.1 Initialization of individuals

Initial population of candidates are traditionally generated in a random way, and

some of them may occasionally satisfy the constraints and requirements, not to

mention optimizing the performance at the same time. When the scale of region or

the size of WSN becomes large, there will be less chance that an initial candidate

has a valid WSN topology. To efficiently tackle this issue, a high ”quality” initial

seed is generated to guarantee the basic requirements on connectivity and coverage.

The LowCost heuristic proposed in [42] is employed to add a valid individual at the

initial phase:

At the beginning, the Coverage is computed for each deployable point m ∈ At

and the heuristic selects mi that with the maximum Coverage as the best location

and a sensor node Su is then placed on mi. The coverage level of the monitoring

points newly covered by Su is updated, and those points with a sufficient coverage

level are removed from the set of sensing area As. This procedure is repeated until

all the monitoring points are k-covered.

Afterwards, LowCost focuses on the connectivity problem. Let Nu be the node

of unconnected nodes U. The algorithm selects a node Nc in the connected sensor

nodes C that is the closest to Nu and computes the new �virtual�position m′ of

Nu by moving it towards Nc as long as the set of monitoring points initially covered

by Nu remains unchanged. If Nu is still unconnected after changing its position,

extra relay nodes are put on the line between Nu and Nc so that Nu and Nc are

connected.

The resulted initial seed is expected to be better than a randomly generated

seed which do not guarantee the coverage and connectivity. However, as discussed

in Chapter 1, the result does not solve the optimization between connectivity and

cost. Besides the individual generated by LowCost heuristic, the rest individuals are

generated based on the constraints on location with various length (size of WSN),

thus the initial population of parents (Parent) are obtained and evaluated.

124


A

B

A’

B’

SN

RN

offsetIcropL

offsetI 'cropL

offsetIcropL

offsetI 'cropL

Figure 4.10: Crossover with different lengths.

A

SN

RN

A’

A

A’

A’’ A’’

MM �' MM �'

Figure 4.11: Mutation with changeable length.

4.4.2 Crossover and mutation

At each generation, the parents are recombined (crossover) and mutated with

different probability. The demonstrations of crossover and mutation with variable

individual lengths are shown in Fig. 4.10 and Fig. 4.11 respectively.

125


4.4.2.1 Crossover

Two candidates (Parent(i) and Parent(j)) are randomly selected from the

population of parents and crossover occurs between them with a chance of Pco.

Note that the length of both parents might be different, and the crop length Lcrop is

limited by the shorter length: Lcrop < min(Mi,Mj). And the offsets can be different

in both parents with constraints: Ioff + Lcrop ≤ M . Afterwards, the two generated

children (A′ and B′) are stored into the population of Children. If crossover does

not happen on the current couple, the two parents are stored into Children directly

for further modification. Similar procedure repeats on other possible pairs of Parent

until Children is filled by individuals.

4.4.2.2 Mutation

Mutation occurs on the individuals of Children. The size of Children(i) isMi which

mutates with a probability of Pmu. Note that the change on Mi will to some extent

increase the risk of constructing an un-connected topology. The new length M ′i is

obtained by randomly selecting a value within the range limited by [Tmin, Tmax],

where Tmin = max(Mi−T, 1), and Tmax = min(Mi+T,Mmax). T is a small integer

value and is equal to 2 in our work. Mmax is the maximum number of nodes that

allowed to be used and is determined by user. If M ′i is less than Mi, M

′i nodes are

randomly picked from Children(i); otherwise, if M ′i > Mi, M

′i −Mi random nodes

are added to Children(i).

Thereafter mutation happens on each node of Children(i) with probability of

Pmu, on the 3D location and the type of node. The movement of Nj ∈ Children(i)

is limited within the sphere of radius dmax centered at Nj .location and the type can

be selected randomly between SN and RN.

4.4.3 Evaluation based on desirability models and constraints

Routes are searched from each node to the BS by using the Dijkstra’s method.

Several further steps are prosecuted to make the algorithm converge faster: A SN is

changed to RN when it does not cover any point m ∈ As; A RN is deleted if it does

not act as a router for other nodes; If two nodes of the same type are located too

126

4.5. Experimental results and analysis

close to each other, one of them are moving apart in a similar way as the mutation

on location.

The desirability values of all the 5 metrics are computed and then attached to

each corresponding individual. Parent and Children are mixed so that all the

individuals from both populations are ranked based on nondominated sorting by

NSGA-II. As a WSN must focus on fulfilling the sensing tasks, the desirability of

coverage Dc is considered as the only constraint among the five objectives. By doing

so, if Dic > Dj

c , the rank of individual i is always higher than individual j no matter

how the other metrics are; otherwise, the ranking is based on all the desirability

values equivalently, the greater a desirability is, the better the corresponding metric

will be. At the end of each generation, Np best individuals are selected and formulate

new Parent for the next generation.

The evolutionary procedure repeats until the maximum generation is reached.

The proposed method is able to provide multiple WSN deployment solutions with

optimized performance from different aspects. As a result, it gives designers

flexibility to observe different optimized deployments and assist them making

deployment decision accordingly.

4.5 Experimental results and analysis

The performance of planning algorithm is evaluated through observing the fitness

value and time efficiency compared with other heuristics, as well as checking the

feasibility in real applications. The performance comparisons are realized with three

comparable state-of-the-art algorithms which have 3D computation ability. All the

algorithms including the proposed one are programmed in C++ and they are run on

a PC with Intel Core i5-760 2.8 GHz CPU so that the results are fairly compared.

4.5.1 The impact of maximum number of generation

As the proposed algorithm is based on evolutionary strategy, the larger the number

of generations, the more outstanding the population will evolve. Therefore, we first

evaluate how the maximum number of generations NUMmaxGen impacts on WSN

performance. The application requirements are shown in Fig. 4.12: The 3D map is

the floor plan of CEI-UPM with a scale of 57 m × 16 m × 3 m and a resolution of

127


1 m. The red point indicates the location of BS (17.52 m, 10.02 m, 1.5 m) and blue

rectangles are the sensing areas As which contain 70 points to be covered.

BS

Target region

Target region

Figure 4.12: Configuration of Scenario CEI-UPM.

In this study, the size of population is Np = 8 for both Parent and Children.

Crossover and mutation possibilities are Pco = 0.1 and Pmu = 0.2 respectively. The

data period for each sensor node is 1 s and the simulation lasts 2400 s in WSNet

simulator. The maximum value of generation NUMmaxGen increases from 10 to

150 with a step of 10, hence there are 15 different NUMmaxGens. The algorithm

runs 5 times for each NUMmaxGen, and as a result 8 × 5 × 15 optimized solutions

are obtained after the simulation. The results are grouped for each NUMmaxGen

and the pareto front is shown in Fig. 4.13, the mean value of each group data

are calculated for each desirability metric, the five metrics construct a plot with five

axes. The area constructed from each group indicates that as NUMmaxGen increases

and the area grows larger, and the overall performance becomes more stable when

NUMmaxGen ≥ 100. Fig. 4.14 indicates that although time consumption fluctuates

at the same NUMmaxGen, consumed time increases approximately linearly with

NUMmaxGen. Therefore, a proper trade-off decision should be made between the

performance and efficiency of computation, and NUMmaxGen = 100 can be selected

for this configuration.

128


maxGen

maxGen

maxGen

Coverage

Cost

Lifetime

Packet latency

Packet drop rate

Figure 4.13: Desirability values vary with NUMmaxGen.

10 20 30 40 50 60 70 80 90 100 110 120 130 140 1500

50

100

150

200

250

300

350

Max number of generation

Tim

e C

onsu

mpt

ion

(s)

Figure 4.14: Time consumption varies with NUMmaxGen.

129


Table 4.3: Features of algorithms for comparison

Algorithm WSN type Solutions RadioObjectivesC Cost L Pl Pd

proposedSN,RN, BS

MultipleRay-tracing(RT)

Y(RT) Y Y Y Y

WMOGASN,RN, BS

SingleRay-tracing(RT)

Y(RT) Y Y Y Y

MOGA SN, BS Single distance Y(dist.) N Y N N

LowCost SN, BS SingleLine-of-Sight(LoS)

Y(LoS) Y N N N

4.5.2 Performance comparison with other heuristics

Two comparable heuristics (LowCost [42] and MOGA [43]) are selected and

programmed in the same platform as the proposed algorithm, so that their

performance are fairly evaluated in exactly the same configuration. Besides, a

weighted multi-objective fitness function (WMOGA) developed in our previous work

by [140] is modified and implemented to evaluate the impacts of NSGA-II on the

final solutions, the formula of the weighted function is:

f = w1DC + w2Dcost + w3DL + w4DPl+ w5DPd

(4.9)

Table 4.3 compares the features of the algorithms, and the proposed method

considers more objectives and more practical modelling for heterogeneous WSN. We

do not analyze the impact of modelling the radio and sensing signal in this thesis,

however it has proved in the previous work, the practical ray-tracing algorithm

outperforms other distance based empirical models. By setting reflection and

diffraction depth as depth = 0 for the ray-tracing engine, only direct paths are

traced in this evaluation. Therefore the algorithms are compare in time efficiency

and optimization performance.

The population of Children is 8 for the first 3 algorithms, and the parent is 8

for the proposed method, 1 for the other two. NUMmaxGen = 150 for the first 3

algorithms. Three scenarios are tested for all the algorithms and the configurations

are: the first scenario is in the CEI-UPM floor plan, As = 70 and BS.location =

(17.52 m, 10.02 m, 1.5 m) (Fig. 4.12); the second scenario is based on the floor plan

of East Lansing which has a scale of 77.4 m × 36.6 m × 3 m and the resolution is

130


1 m. As = 232 and BS.location = (23.43 m, 13.60 m, 1.5 m) (Fig. 4.15); outdoor

region of Madrid city is the third scenario. Its scale is 233.36 m× 297 m× 73.67 m

and resolution is 3 m. As = 310 and BS.location = (107.8 m, 60.1 m, 3 m) (Fig.

4.16).

BS

Target region

Target region

Target region

Figure 4.15: Scenario East Lansing.

The five desirability metrics used by the proposed method are also computed

based on the topologies generated by the other three algorithms, the solutions are

compared according to those values as indicated in Table 4.4, 4.5 and 4.6 respectively.

Red color marks the best value and blue color marks the worst. The performance

of MOGA is limited by using a constant WSN size and unilateral objectives.

WMOGA considers all the objectives without the ability of providing multiple

solutions simultaneously, which explains why this method obtains good optimization

performance. As expected, most of the best results are obtained by the proposed

algorithm, allowing designers making decisions from different aspects to construct

a reliable topology. The computation time of all the algorithms increases as the

scale of scenario grows. LowCost is the most efficient with deterministic heuristic

while it obtains worst performance on some objectives at each scenario. Especially

when there are significant obstacles between the unconnected and connected part of

131


BS

Target region

Target region

Figure 4.16: Scenario Madrid.

a WSN, more nodes are placed across the obstacles instead of placing around them

to optimize the cost. The proposed method provides 8 different multi-objective

optimized solutions simultaneously and leads to the highest time consumption at

each scenario. However if divided by 8, the average time consumption per solution

is at least 45% better than MOGA and at least 60% better than WMOGA. As a

result, even by integrating external WSNet simulation in the loop of evolutionary

strategy and with changeable size of WSN, the proposed method is still much more

efficient than MOGA and WMOGA.

132

4.6. Conclusion

Table 4.4: Results comparison for Scenario CEI-UPM.

Algorithm Time (s)ObjectivesDC DCost DL DPl

DPd

Proposed 255.09

1 0.84 0.24 0.99 0.961 0.835 0.29 0.99 0.861 0.83 0.25 0.99 11 0.84 0.22 0.99 0.921 0.845 0.24 0.99 0.881 0.84 0.22 0.99 0.991 0.85 0.26 0.99 0.981 0.85 0.25 0.99 1

WMOGA 456.64 1 0.835 0.26 0.99 0.98

MOGA 67.66 1 0.82 0.28 0.98 0.9

LowCost 1.3 1 0.82 0.21 0.99 0.96

Table 4.5: Results comparison for Scenario East Lansing.


DPd

Proposed 799.32

1 0.805 0.26 0.99 0.911 0.81 0.24 0.99 0.971 0.83 0.21 0.99 0.941 0.81 0.26 0.99 0.921 0.815 0.30 0.99 0.961 0.815 0.30 0.99 0.871 0.82 0.21 0.99 0.891 0.815 0.32 0.99 1

WMOGA 581.03 1 0.815 0.36 0.99 1

MOGA 177.14 1 0.82 0.28 0.99 0.94

LowCost 3.64 1 0.82 0.19 0.99 0.96

4.6 Conclusion

A planning algorithm is proposed in this thesis, by taking the advantage of proposed

ray-tracing scheme for both radio and sensing signal propagation, the modeling on

coverage and connectivity turns out to be accurate. Moreover, we also consider

the important impact of lifetime and link quality on the WSN, the optimization

is more complete compared with other works. This algorithm is suitable for both

outdoor and indoor environment with the ability to consider deployable area and

133


Table 4.6: Results comparison for Scenario Madrid.


DPd

Proposed 4048.57

1 0.83 0.19 0.99 0.951 0.83 0.18 0.99 0.971 0.83 0.20 0.99 0.951 0.86 0.22 0.99 0.911 0.85 0.21 0.99 11 0.86 0.23 0.99 0.991 0.83 0.20 0.99 0.941 0.865 0.20 0.99 1

WMOGA 1236.32 0.99 0.755 0.18 0.99 1

MOGA 1581.14 1 0.58 0.17 0.99 0.99

LowCost 35.29 1 0.58 0.05 0.99 0.83

forbidden area. The scalability and probabilistic model of sensing signal are also

very challenging topic in this area, and should be tackled in the future.

134

Chapter 5

iMOST: an Intelligent

Multi-objective Optimization

Sensor network planning Tool

iMOST is developed by integrating the introduced algorithms, to assist WSN

designers efficiently planning reliable WSNs for different configurations. The

abbreviated name iMOST stands for an Intelligent Multi-objective Optimization

Sensor network planning Tool. As mentioned in Section 1.4 of Chapter 1, the

structure of iMOST is illustrated in Fig. 1.9, which is composed of an user friendly

interface and three core functional modules: Image Processing Module (IPM),

Ray-tracing Propagation Module (RPM) and Node Placement Module (NPM).

iMOST contributes on: (1) Efficient and automatic 3D database reconstruction

and fast 3D objects design for both indoor and outdoor environments; (2) It

provides multiple multi-objective optimized 3D deployment solutions and allows

users to configure the network properties, hence it can adapt to various WSN

applications; (3) Deployment solutions in the 3D space and the corresponding

evaluated performance are visually presented to users; and (4) The Node Placement

Module of iMOST is available online as well as the source code of the other

two rebuilt heuristics. Therefore WSN designers will benefit from this tool on

efficiently constructing environment database, practically and efficiently planning

reliable WSNs for both outdoor and indoor applications. With the open source

codes, they are also able to compare their algorithms with ours to contribute this

academic field.

The current version of iMOST works on Win7 operating system and its

mainframe is programmed by using MFC of Visual Studio. The details of important

features of iMOST are introduced in this chapter.

Chapter 5. iMOST: an Intelligent Multi-objective Optimization Sensornetwork planning Tool

The mainframe of the user interface is shown in Fig. 5.1. It is consist of

Menu bar, Tool bar where functions can be realized based on user demands and

Demonstration area where 3D view of scenario and WSN evaluation result can be

demonstrated. The meaning of each icon and the corresponding function is indicated

in the figure as well. The IPM is called by the NewMap menu in File option (ref.

to sec. 5.1.1), map will be loaded to the Demonstration area by Open. The tool

bar provides 7 options including: manually place nodes on the loaded map, view the

performance of WSN, define sensing region, construct WSN topology based on the

planning algorithm, reset the scenario and navigate the scenario and deployment in

3D.

5.1 Menu bar

5.1.1 Image Processing Module

By selecting NewMap menu under File, the IPM is triggered and automatic

environment reconstruction algorithm is run for either indoor or outdoor scenario

based on the procedure introduced in Chapter 2. As shown in Fig. 5.2, user should

first define the directory of tested images, the folder of trained database and the

folder of output results. The automatic 3D environment reconstruction method

is launched after user finishing configuration, images are loaded and objects are

recognized automatically. This procedure allows users freely preparing their own

environment and reconstructing the environment via the trained database, so that

the 3D database is constructed efficiently and accurately without spending many

human efforts or high cost compared with conventional methods.

5.1.2 Environment property configuration

Once the 3D environment database is constructed or there is any 3D map available

in ’txt’, ’kml’ or ’3DS’ file with the polygonal description format, the map can be

imported to the tool and demonstrated in the visual interface by clicking Open

menu. During the importing of map, the Environment Property dialog is pop up,

which is shown in Fig. 5.3. It provides opportunity for users setting the environment

properties including the resolution and scale ratio compared with the real world size.

136

5.1. Menu bar

Sensing region defineManual node deployment mode Planning algorithm Navigation of

scenario

View candidate result Finish configuration Clean the deployment

Scenario and result demonstration

IPM: Automatic environment reconstruction

Import available 3D map: ‘txt’, ‘kml’, ‘3DS’ format

Figure 5.1: The mainframe of the planning tool.

Figure 5.2: User command on constructing new map.

137


Figure 5.3: Environment property setting dialog.

The resolution also determines how many points the ray-tracing engine will calculate

for, and the number of possible locations of nodes as well. Therefore, resolution will

impact on the computation time of the algorithms, whereas the scale ratio only

affects the real scale of the scenario and topology of nodes.

5.2 Toolbar

5.2.1 Node deployment

Beyond the automatic topology generation by the proposed planning algorithm,

nodes can be placed by users for other purpose such as to evaluate the performance

of a designed topology. The node deployment button allows users placing nodes

on determined location, and a Node Property dialog ( see Fig. 5.4) will pop out

each time the left mouse button clicks on the valid region, to allow configuring

the property of the placed node on type: SN, RN and BS, Tx power in dBm,

radio sensitivity Rxs in dBm and if it is a sensor node the sense range can be

set in meters. As expected each node may have different communication ability

and sensing ability defined by users. Different types of nodes are distinguished by

different colors and point sizes in this work, for example red color point with the

largest size represents a BS, the point with purple color and middle size is a

RN and yellow color point with the smallest size indicates a SN.

5.2.2 Network Planning Module

The NPM requires users make 3 steps to configure the network:

138

5.2. Toolbar

Figure 5.4: Node property configuration dialog.

• Define sensing area As. By pressing , the function of definning sensing

area is triggered. User click the left mouse button on the scenario to mark

the border corners of sensing region as shown in Fig. 4.12. In the current

version, each region is represented by four vertexes to construct a quadrangle.

However, the shape of region is not limited in the real world, this function can

be extended by allowing more vertexes for a region.

• Node pre-deployment. Several nodes can be pre-deployed by users by

pressing . This tool supports semi-auto and auto planning of WSN, where

semi-auto means that user pre-define parts of WSN on specified locations with

determined properties, whereas only the BS is determined manually in auto

planning mode.

• Terminate user configuration and launch the planning algorithm.

Once the button is pressed, the configuration is done and user can launch

the proposed planning algorithm by pressing . At the end, multiple

multi-objective optimized solutions can be viewed visually by clicking on the

icon and then selecting the individual ID from the dialog shown in Fig.

5.5. See Fig. 5.6 the results of individual ID = 4 as an example, with the

topology and the value of each objective demonstrated.

139


Figure 5.5: Node configuration dialog.

Coverage Cost Lifetime Packet latency Packet drop rate

1.00 0.855 0.22 0.98 1

Figure 5.6: Generated topology of individual with ID=4.

5.2.3 Ray-tracing Propagation Module

The RPM can not be visually seen from the mainframe, however it is embedded

in the function behind the icons of view results and the planning algorithm

. Whenever the radio propagation or sensing coverage is needed, they are all

computed via RPM so that the accuracy and reliability of the estimated result is

guaranteed.

5.2.4 3D navigation

The tool offers the capability to navigate the loaded 3D scenario by pressing .

The scenario or deployment of WSN can be viewed with different angles of view

and zooming scales. The zooming function is realized by rolling the scroll wheel of

the mouse forward or backward. The view point is adjusted through pressing the

right mouse button and moving the mouse at the same time. The angle of views

140

5.2. Toolbar

Figure 5.7: 3D navigation for outdoor scenario.

are changed by pressing the left button and moving the mouse simultaneously. Fig.

5.7 and Fig. 5.8 demonstrate the 3D navigation at outdoor and indoor environment

respectively. The generated topology, the radio rays computed from RPM and the

evaluated network performance can all be shown in 3D view. Moreover, additional

objects in the scenario can be eliminated according to the user requirement. By

doing so, they don’t have any impacts on radio propagation and sensing signal

modelling.

141


Figure 5.8: 3D navigation for indoor scenario.

5.3 Conclusion

iMOST is a tool developed to assist designers planning WSN topology with

optimized performance. It features with high efficiency, flexibility and reliability

with low cost by integrating the proposed IPM, RPM and NPM by this work.

With the ability of automatic 3D outdoor and indoor environment reconstruction,

it significantly reduces economic cost, human efforts and time that spent on this

crucial but non-planning issue. This tool also explores the feasibility of a new

direction of image processing application. User-friendly interface allows designers to

configure the scenario and WSN properties conviniently. It visually demonstrates

the deployment solutions which not only satisfiy the design requirements but also

overall optimize the WSN performance by NPM on sensing coverage, connectivity,

hardware cost, network longevity, packet latency and packet drop rate. The node

placement module and the rebuilt heuristics are accessible through internet so as to

benefit other WSN designers from different aspects. As part of the work, the WSN

deployment solutions will be evaluated under real measurements.

142

Chapter 6

Real measurements and results

analysis

Two environment monitoring demonstrations were set up to validate the

performance of the proposed planning algorithm and the developed planning

tool. The overall methodology, from 3D environment reconstruction and user

configuration to radio propagation, topology generation and performance estimation,

is employed for both demonstrations.

In both tests, ’Cookie’ is equipped with ZigBee communication protocol layer,

environment sensors that are able to sense temperature, light and humidity, and

external antennas made by EAD, see Fig. 6.1. The BKR2400 antenna is 1/2 wave

dipole with 2 dBi peak gain, it has linear polarisation with omni-directional radiation

pattern at horizontal plane. Lithium-based batteries have been used to supply the

energy to the modular nodes during the WSN deployments and experimental tests,

providing up to 500 mAh which covers the power consumption requirements of the

devices. Moreover, they can be charged by using the power supply layer of the

Cookie architecture, so that the autonomy of the nodes is enhanced.

6.1 Aggregation mechanism of measured data

Besides gathering the ambient data, the most important parameters that should

be observed are RSS value of data packets, neighborhood table of each node,

routing table with the BS as destination, battery level and packet delivery states.

The request of those observations is implemented by means of using the HW-SW

co-design platform proposed by [141], which is a framework based on libraries and

controllers that allows designers realizing applications by programming in C code

Chapter 6. Real measurements and results analysis

Figure 6.1: BKR2400 antenna.

and compiling to generate the bitstreams for the microcontroller of sensor node. The

mechanism of aggregating the aforementioned data is described as follows:

1. RSS. Whenever a node Ni receives packets from other nodes, it records the

RSS in dBm, and noted as

RSS(Nj , tk)

Where Nj is the source of a packet, tk ∈ [0, tp] is the arrival time stamp of the

packet, tp is the overall testing time of WSN.

2. Neighborhood table of each node. The neighborhood table Ti of Ni

contains the IDs of neighbors and corresponding RSS records. The format of

Ti is expressed as following and for simplicity in real application, 10 samples

of RSSs are cached for each neighbor of Ni.

Ti =

⟨ Nj1 RSSI(Nj1 , t0), RSSI(Nj1 , t1), · · · , RSSI(Nj1 , tp)

Nj2 RSSI(Nj2 , t′0), RSSI(Nj2 , t

′1), · · · , RSSI(Nj2 , t

′p)

. . .. . .

⟩

3. Routing table of the network. It records for the whole WSN the routes

from each node to BS. Each node maintain the next hop information, so that

at the base station, routing table R of the WSN is constructed:

144

6.2. Application interface

Table 6.1: Routing table format.

Source Destination Next hop

Ni BS Nk

Nj BS Nv

. . . BS. . .

4. Power consumption (Battery level). This observation is to estimate the

lifetime of the network. By observing battery level, the battery status of each

node can be evaluated. Each node records its battery status BL at every

period tB and reports to the BS every n · tB. By doing so, the battery level

table B of the whole network is expressed as following:

B =

⟨ Nj1 BL(Nj1 , tB), BL(Nj1 , 2tB), · · · , BL(Nj1 , ntB)

Nj2 BL(Nj2 , t′B), BL(Nj2 , 2t

′B), · · · , BL(Nj2 , nt

′B)

. . .. . .

⟩

5. Format of data packet. The BS gathers sensed data from sensors. Each

SN periodically (Td) sends data packet to BS, the data packet contains the

following information:

Table 6.2: Packet format.

Source ID Arrival time stamp Sequence Sensed data TX time stamp

Once the test is terminated, the packet loss rate can be analyzed based on the

continuity of packet sequence for each node, and packet delay of the WSN is

computed from the differences between TX time stamp and arrival time stamp.

Moreover, environmental monitoring data are obtained as well to prove that

those solutions can satisfy the application requirements.

6.2 Application interface

There is usually a user application of observing data and maintain the functions of

nodes in any type of WSN applications. We also developed a user interface for our

145


demonstrations to allow simple operations from users. The application interface is

programmed by using JAVA with all the aforementioned function to fetch all the

required information. It works in cooperation with the BS and Fig. 6.2 shows the

structure of the interface.

Figure 6.2: Application interface.

• Data packet is shown once received by BS. As can be seen, the source ID,

TX time, temperature(T), humidity (H), light (L), battery level and sequence

number are included. Arrival time stamp is added at the end.

• Routing table can be generated according to user’s command, which is realized

by clicking ’Generate Routing Table’ button.

• User can select node ID and click the ’Neighbor Table’ to observe the

neighborhood table of a node.

146

6.3. Indoor measurements

• Time is synchronized by clicking ’Configure Time’.

• All the information is saved to ’.txt’ file once ’Save Current Information’ button

is pressed.

After preparing all the aforementioned hardware devices, aggregation mechanism

and user interface software are used to monitor the WSN. Two real deployments

are launched for an indoor and an outdoor environment to monitor temperature,

humidity and light level and validate the planned topology and performance

estimated by iMOST.

6.3 Indoor measurements

The first test is realized in the indoor environment of CEI-UPM. The 3D

indoor environment database is constructed based on the Chapter. 2, where the

reconstructed results are analyzed in details. The scanned map is shown in Fig.

6.3(a). The reconstructed result (Fig. 6.3(b)) has 88 planes accurately constructed

and 2 edges misclassified. Beyond the automatically provided environmental

database, we manually import other 3DS models (Fig. 6.3(c)) to show the presence

of office desks and consider their impacts on the topology planning. Environment

database is loaded to iMOST, resolution is equal to 1.0 m and scale ratio is 1.0.

User requirements on sensing regions are demonstrated in Fig. 6.4, As = 90

and BS.location = (18.73 m, 12.00 m, 1.5 m). Nodes are set with TX power as

−12 dBm, RX sensitivity as −98 dBm, Rsense = 3 m.

A population of planed topologies are generated by iMOST based on the user

configuration with optimized coverage, cost, lifetime, packet latency and packet

drop rate. In this work, we select two candidates for real deployments, which

have topologies indicated in Fig. 6.5(a) and Fig. 6.6(a). All the nodes are

placed approximately at the locations indicated by the planned solutions and the

constructed topologies are shown in Fig. 6.5(b) and Fig. 6.6(b) respectively. The

routing from N5 to BS in the planned solution is via N4 whereas it routes though

N8 in the real deployments. Except N5, the remaining nodes have the same next

hop as indicated by the iMOST solutions.

147


(a) Original scanned floor plan of CEI-UPM

(b) Automatic reconstructed result

(c) 3DS imported models

(c) 3D view of the scenario with furniture

Figure 6.3: Indoor modelling by using iMOST: automatic 3D reconstruction+3DSmodels.

The evaluated performance of the two candidate solutions are shown in Table

.6.3, topology 1 performs better than topology 2 in terms of cost, lifetime and data

drop rate.

Table 6.4 shows the measured data of topology 1 in details. The detected

neighbors and corresponding RSS values are shown (the first sub-row) for each

node (N1-N17) in the scenario. All the detected RSS values are computed by

averaging the fetched samples along different time, and they are compared with

148


Figure 6.4: User requirement over the indoor test.

(a) iMOST solution

(b) Real topology

1 2 3 4 5 6 7

98 10

11 12 13 14 15 16 17

BS

Figure 6.5: Topology comparison 1: (a) One of the eight solutions generated byiMOST for CEI-UPM. (b) The topology of real deployment.

149


(a) iMOST solution

(b) Real topology

1 2 3 4 5 6 7

98 10

11 12 13 14 15 16 17

18

BS

Figure 6.6: Topology comparison 2: (a) Another solution generated by iMOST forCEI-UPM. (b) The topology of the corresponding real deployment.

Table 6.3: Evaluated performance of the two candidates.

Topology solutionObjectivesDC DCost DL DPl

DPd

1 1 0.83 0.432 0.99 0.992 1 0.825 0.428 0.99 0.95

the simulation results, by setting the ray-tracing engine with maximum reflection

depth depth = 0(the second sub-row) and depth = 3 (the third sub-row). The

number of traced rays grows as depth increases, however the results will be more

accurate as more multi-path effects are considered.

As can be seen, there are some errors in discovered neighbors of some nodes (e.

g.N2 and N14), and the errors are marked by red color. Assuming that N(u) is the

set of actual neighbors of a node u, and N ′(u) the set of neighbors known to u (i.e.

whose identifier is present in its neighborhood table). Neighborhood accuracy Accnb

is the average value of the accumulation for all the nodes in the WSN the proportion

150


of actual neighbors of node that have been indeed detected. It is formulated as (6.1):

Accnb =

∑Mu=1

|N(u)⋂

N ′(u)||N ′(u)|

M× 100%. (6.1)

Accordingly, we have the accuracy of neighborhood Accnb = 88% in this case

study. The main reasons of this error are from the embedded HELLO protocol of the

ZigBee layer. As N2 cannot detect N7 and N8 while N7 and N8 is able to discover

N2 as their neighbor, the similar reason can be applied for the neighborhood of

N5,6,10,12,13,14,16.

Fig. 6.7 demonstrates graphically RSS values of the measured data and

simulation results when depth = 0 and depth = 3, by aligning the actual neighbors

for each node (from N1 to N17). Table 6.4 shows the mean error (ME) and standard

deviation error (STD) for those two simulations. When depth = 0, ME0 = 4.29 dB

and STD0 = 5.06 dB; If depth = 3, ME3 = 3.80 dB and STD3 = 3.61 dB, which as

expected, performs better (11% reduction inME and 28.6% reduction in STD) than

the former case. Therefore WSN designer should make a trade-off decision between

the time consumption and accuracy, if the planning strategy can be realized without

time constraint, this work suggests users configure ray tracing engine with depth > 0

to improve the accuracy of topology estimation.

This scenario is tested on 22nd and 23rd November, 2013. Sensors detect

environment and send data packet every one minute, the battery status is reported

from each node to BS every one minute as well. This traffic load mechanism aims to

speedup the battery consumption with a ratio of 120 times faster than the simulation

period (100 days). Therefore network longevity is predicted as 8.64 hours and we

should change battery since that moment. The curves of battery consumption shown

in Fig. 6.8 are the measured remaining energy of N1,8,14 varied along working time

of WSN. According to the test N1 has the lowest lifetime around 8.3 hours, which

obtains 96.1% of match with the predicted value the estimated result and indicates

a good performance of the proposed work on lifetime modelling in iMOST.

Fig. 6.9 compares packet delivery status in a period of 30 minutes for topology 1.

N7 has the highest packet loss rate (20%) because it needs the maximum number of

hops to reach BS compared with other nodes. The average packet latency is around

3.5 s therefore Dpl ≈ 0.97 which is slightly less (3%) than the estimated result.

151


Table

6.4:Neig

hborhoodtable

andRSScompariso

nsbetw

eenrea

lmeasurem

entandsim

ulatio

nresu

lts:indoorscen

ario

.

NodeID

NeighborID

and

Avera

geRSSs(d

Bm)

Accnb

RSS

ME,STD

1BS(-6

9.00),2

(-55.00),3

(-65.00),4

(-70.50),5

(-78.00),6

(-79.00),7

(-84.00),8

(-63.00),9

(-65.00),1

0(-8

3.00)

BS(-6

8.98),2

(-54.25),3

(-64.55),4

(-68.82),5

(-76.35),6

(-78.43),7

(-80.68),8

(-65.32),9

(-69.40),1

0(-7

2.00)

BS(-6

8.85),2

(-54.27),3

(-64.51),4

(-70.05),5

(-86.31),6

(-88.52),7

(-90.52),8

(-65.68),9

(-69.31),1

0(-8

2.08)

2BS(-7

4.00),1

(-55.00),3

(-48.00),4

(-65.67),5

(-79.00),6

(-73.00),9

(-65.00),1

0(-7

0.00)

BS(-7

1.10),1

(-54.25),3

(-60.28),4

(-65.88),5

(-74.58),6

(-76.82),7

(-79.21),8

(-64.57),9

(-67.37),1

0(-6

9.99)

BS(-7

2.85),1

(-55.67),3

(-59.62),4

(-66.11),5

(-85.83),6

(-86.68),7

(-89.21),8

(-64.74),9

(-67.42),1

0(-6

9.95)

3BS(-7

2.00),1

(-63.50),2

(-62.00),4

(-61.75),5

(-75.00),6

(-69.50),7

(-78.25),8

(-71.00),9

(-63.50),1

0(-6

9.00)

BS(-7

4.12),1

(-64.55),2

(-60.28),4

(-57.78),5

(-71.35),6

(-74.02),7

(-76.80),8

(-67.99),9

(-63.72),1

0(-6

7.37)

BS(-7

4.94),1

(-64.01),2

(-60.23),4

(-57.84),5

(-81.80),6

(-84.00),7

(-76.8),8

(-88.05),9

(-53.00),1

0(-6

8.40)

4BS(-7

0.00),1

(-74.00),2

(-69.00),3

(-61.00),5

(-66.00),6

(-74.00),7

(-77.00),8

(-73.00),9

(-63.00),1

0(-6

4.50)

BS(-7

6.27),1

(-68.82),2

(-65.88),3

(-57.78),5

(-68.13),6

(-71.35),7

(-74.56),8

(-69.24),9

(-66.81),1

0(-6

3.82)

BS(-7

4.89),1

(-70.05),2

(-66.23),3

(-57.84),5

(-68.22),6

(-71.15),7

(-75.07),8

(-69.15),9

(-65.56),1

0(-6

3.96)

5BS(-7

8.50),1

(-86.00),3

(-76.00),4

(-62.25),7

(-55.00),8

(-81.00),9

(-72.33),1

0(-7

3.00)

BS(-8

0.98),1

(-76.35),2

(-74.58),3

(-71.35),4

(-68.13),6

(-57.78),7

(-65.99),8

(-76.26),9

(-72.12),1

0(-6

9.43)

BS(-8

0.32),1

(-86.63),2

(-84.33),3

(-72.17),4

(-68.22),6

(-57.71),7

(-60.45),8

(-76.52),9

(-72.06),1

0(-6

9.66)

6BS(-7

2.00),1

(-83.00),3

(-72.00),4

(-74.30),5

(-55.00),7

(-45.00),9

(-74.00),1

0(-7

0.00)

BS(-8

1.85),1

(-78.43),2

(-76.82),3

(-74.02),4

(-71.35),5

(-57.78),7

(-60.58),8

(-77.55),9

(-75.28),1

0(-7

2.25)

BS(-7

9.33),1

(-78.52),2

(-77.71),3

(-74.49),4

(-71.15),5

(-57.71),7

(-56.86),8

(-77.31),9

(-75.27),1

0(-7

2.14)

7BS(-4

3.00),1

(-91.00),2

(-87.60),3

(-80.33),4

(-75.00),5

(-54.25),6

(-44.00),8

(-79.50),9

(-76.00),1

0(-7

1.00)

BS(-8

4.42),1

(-80.68),2

(-79.21),3

(-76.80),4

(-74.56),5

(-65.99),6

(-60.58),8

(-79.89),9

(-78.03),1

0(-7

4.75)

BS(-5

4.30),1

(-80.33),2

(-80.87),3

(-76.80),4

(-75.05),5

(-60.45),6

(-56.85),8

(-79.69),9

(-76.39),1

0(-7

4.77)

8BS(-6

8.00),1

(-60.00),2

(-65.00),3

(-73.00),4

(-77.00),5

(-76.00),6

(-71.00),7

(-76.50),9

(-52.00),1

0(-7

6.00)

BS(-6

9.23),1

(-65.32),2

(-64.57),3

(-67.99),4

(-69.24),5

(-76.26),6

(-77.55),7

(-79.89),9

(-62.21),1

0(-6

8.07)

BS(-6

9.04),1

(-65.68),2

(-63.70),3

(-68.37),4

(-69.10),5

(-76.22),6

(-77.84),7

(-79.69),9

(-62.37),1

0(-6

8.09)

ME

0=

4.29

9BS(-8

3.00),1

(-67.50),2

(-56.00),3

(-58.00),4

(-66.00),5

(-75.00),6

(-75.00),7

(-76.00),8

(-65.50),1

0(-6

4.50)

STD

0=

5.06

BS(-7

3.14),1

(-69.40),2

(-67.37),3

(-63.72),4

(-66.81),5

(-72.12),6

(-75.28),7

(-78.03),8

(-62.21),1

0(-6

0.28)

88%

BS(-8

3.58),1

(-68.95),2

(-66.33),3

(-64.00),4

(-66.76),5

(-72.06),6

(-75.27),7

(-78.01),8

(-62.74),1

0(-6

0.25)

10

BS(-5

6.00),1

(-72.00),2

(-64.00),3

(-66.00),5

(-64.00),6

(-72.50),7

(-69.00),8

(-69.00),9

(-56.00)

BS(-7

5.80),1

(-72.00),2

(-54.25),3

(-67.37),4

(-63.82),5

(-69.43),6

(-72.25),7

(-74.75),8

(-68.07),9

(-60.28)

ME

3=

3.80

BS(-6

8.89),1

(-72.03),2

(-69.95),3

(-66.96),4

(-63.85),5

(-68.99),6

(-72.15),7

(-74.77),8

(-68.72),9

(-59.58)

STD

3=

3.61

11

12(-5

9.00),1

3(-6

7.00),1

4(-6

4.00),1

5(-7

1.50),1

6(-7

3.00),1

7(-7

9.00)

BS(-7

5.34),1

2(-5

4.25),1

3(-6

5.99),1

4(-6

8.13),1

5(-7

3.25),1

6(-7

5.59),1

7(-7

7.67)

BS(-7

4.90),1

2(-5

4.27),1

3(-6

7.00),1

4(-6

4.76),1

5(-7

4.42),1

6(-7

5.60),1

7(-7

8.90)

12

11(-5

9.00),1

3(-6

0.00),1

5(-8

1.00),1

6(-7

5.00),1

7(-8

1.00)

BS(-7

3.75),1

1(-5

4.25),1

3(-6

2.41),1

4(-6

5.24),1

5(-7

1.26),1

6(-7

3.81),1

7(-7

6.05)

BS(-7

3.30),1

1(-5

4.27),1

3(-6

2.32),1

4(-6

2.95),1

5(-7

1.71),1

6(-7

4.72),1

7(-7

7.28)

13

11(-6

4.00),1

2(-6

2.00),1

4(-5

6.00),1

5(-6

9.00),1

6(-6

5.00),1

7(-6

9.00)

BS(-7

2.62),1

1(-6

5.99),1

2(-6

2.41),1

4(-5

6.87),1

5(-6

6.37),1

6(-6

9.85),1

7(-7

2.68)

BS(-7

3.12),1

1(-6

6.24),1

2(-6

2.61).1

4(-5

5.71),1

5(-6

6.10),1

6(-6

9.49),1

7(-7

3.04)

14

11(-6

4.00),1

3(-5

6.00),1

5(-5

4.00),1

7(-7

2.00)

BS(-6

9.82),1

1(-6

8.13),1

2(-6

5.24),1

3(-5

6.87),1

5(-6

3.19),1

6(-6

7.38),1

7(-7

0.60)

BS(-6

9.54),1

1(-6

6.30),1

2(-6

5.30),1

3(-5

6.61),1

5(-5

9.55),1

6(-6

7.81),1

7(-6

9.74)

15

BS(-7

6.00),1

1(-7

5.00),1

2(-8

1.00),1

3(-6

9.00),1

4(-5

4.00),1

6(-5

8.00),1

7(-6

7.00)

BS(-6

7.94),1

1(-7

4.00),1

2(-7

2.01),1

3(-6

7.12),1

4(-6

3.94),1

6(-5

7.78),1

7(-6

4.55)

BS(-6

7.83),1

1(-7

7.01),1

2(-7

1.71),1

3(-6

9.27),1

4(-6

1.73),1

6(-5

7.67),1

7(-6

4.48)

16

11(-7

3.00),1

2(-7

5.00),1

3(-6

4.00),1

5(-5

6.00),1

7(-5

5.00)

BS(-6

7.03),1

1(-7

6.33),1

2(-7

4.56),1

3(-7

0.60),1

4(-6

8.13),1

5(-5

7.78),1

7(-5

7.78)

BS(-6

6.94),1

1(-7

6.14),1

2(-7

4.72),1

3(-6

9.99),1

4(-6

3.97),1

5(-5

7.10),1

7(-5

7.52)

17

BS(-7

0.00),1

1(-7

7.00),1

2(-8

0.00),1

3(-7

1.00),1

4(-7

2.00),1

5(-6

6.00),1

6(-5

5.00)

BS(-6

9.31),1

1(-7

8.42),1

2(-7

6.80),1

3(-7

3.43),1

4(-7

1.35),1

5(-6

4.55),1

6(-5

7.78)

BS(-7

0.32),1

1(-7

8.89),1

2(-7

7.29),1

3(-7

4.32),1

4(-7

2.24),1

5(-6

4.46),1

6(-5

8.17)

152


0 20 40 60 80 100 120 140-100

-90

-80

-70

-60

-50

-40

-30

Measured value index aligned in the table

RSS

(dB

m)

Real measurementSimulation result depth=0Simulation result depth=3

Figure 6.7: RSS comparison between real measurement and simulation result withdepth = 0 and depth = 3.

0 1 2 3 4 5 6 7 8 90

10

20

30

40

50

60

70

80

90

100

Working time (hours)

Rem

aini

ng e

nerg

y (%

)

node 1node 8node 14

Figure 6.8: Remaining energy of N1, N8 and N14 along the working time of WSN.

The packet drop rate is calculated as the proportion between lost data packets and

total number of packets that have been sent in the WSN. According to the observed

153


packet sequence number for each node, there is 10% of packet loss in this case, thus

the desirability metric DPd= 90%. Compared with the estimated result (99%), the

error is approximately 9%.

1 2 3 4 5 6 7 8 9 10111213141516 17

10

15

20

25

Node ID

Num

ber o

f pac

kets

Arrived packetsSent packets

Figure 6.9: Comparing the number of arrived packet and the number of sent packetfor each node in the deployment.

The three nodes: N1,8,14 are the bottlenecks of this deployment. Although N1

has less children than N8 and same as N14, the accumulated dropped packets hoping

through N1 are less than that via N8 and N14. As a result, N1 should forward more

data packets to BS and a faster exhaustion on the battery energy occurs, which also

reveals why (see Fig. 6.8) the lifetime of N1 is shorter than other nodes.

The three nodes N1,8,14 are the bottleneck of the deployment, although N1 has

less children than N8 and same as N14, the accumulated packet loss rate hoping

through N1 is less than that via N8 and N14. As a result, N1 should forward more

154

6.4. Outdoor measurements

data packets to BS and a faster exhaustion on the battery energy occurs, which also

reveals why (see Fig. 6.8) the lifetime of N1 is shorter than other nodes.

Sensed data are gathered at the coordinator and the environment variation is

recorded for each sensor. Fig. 6.10 shows the variation of temperature, humidity

and light in the room where N4 is placed, and the observation time is from 16 : 00

to 21 : 00 on 22nd Nov, 2013. The average temperature is 29.25◦ C, humidity is

47.13 without significant changes along time. While the light level suddenly reduces

from 90.35% (3700) to 0% (0) at 20:30, it indicates that the light is turned off and

people may have left room after this moment.

16 17 18 19 20 210

10

20

30

40

50

Time of the day (from 16hr~21hr)

Mea

sure

d le

vel

TemperatureHumidityLight (x0.01)

Figure 6.10: The sensed data of N4.

6.4 Outdoor measurements

The second test is realized to monitor the outdoor environment at the parking lot of

UPM. The bird’s view of the environment is taken from Google Maps, as shown in

Fig. 6.11(a), the automatic image understanding algorithm, using Madrid training

database, provides the recognition result shown in Fig. 6.11(b) which achieves 93.3%

155


of accuracy compared with the ground truth. It is expected to assigned uniform

(a) Original image (b) Image recognition result

(c) 3D view of the reconstructed result by adding terrain information from Google Earth

Figure 6.11: Outdoor modelling by using iMOST: automatic 3Dreconstruction+terrain information.

height to the building, while in this case we register the terrain information of this

area is available, and therefore it was used to enhance the performance. The 3D

view of the reconstructed model is shown in Fig. 6.11(c). This 3D environment

modelling took 6.83 minutes to reconstruct a region of 233.36 m× 297 m× 73.67 m.

156


Table 6.5: Evaluated performance of the selected candidate.

Objectives

DC DCost DL DPlDPd

1 0.91 0.155 0.99 0.98

User configuration on sensing area for this test is As = 99 and TX power for

all the nodes are set as 4 dBm which is the maximum communication ability of the

antenna. Nodes are set with RX sensitivity as −98 dBm, Rsense = 8 m. iMOST

generates topology (Fig. 6.12(a)) accordingly, and the candidate, which with the

lowest cost and competitive desirability values of other metrics, is selected in this

case. The predicted performance is shown in Table. 6.5. By placing 9 nodes in

target locations, the real topology is shown in Fig. 6.12(b). N1 ∼ N4 are, as

estimated, directly connected with BS, N7 routes via N4 rather than through N6 in

the solution.

Table 6.6 shows the measured data of the deployed topology in details. All the

detected RSS values are computed by averaging the fetched samples along different

time, and they are compared with the simulation results with depth = 3 (the second

sub-row). N5 and N8 can not discover each other in the real deployment while

the simulated result shows connections between them, and therefore Accnb = 98%.

Fig. 6.13 demonstrates graphically RSS values of the measured data and simulation

results when depth = 3, by aligning the actual neighbors for each node (from N1 to

N9). Table 6.6 indicates that in this case ME3 = 2.45 dB and STD3 = 2.45 dB.

As the outdoor environment is an open area without significant obstacles between

nodes, this performance is slightly better than indoor performance. Both indoor

and outdoor neighborhood results indicate that the proposed modelling methods on

radio propagation and link establishment are practical and reliable to be applied in

real deployments.

This scenario is tested on 23rd November, 2013. Sensing period and battery

period are one minute. The estimated network longevity is predicted as 1.86 hours

with the maximum transmission power. The curves of battery consumptions shown

in Fig. 6.14 are the measured remaining energy of N4,5,2 varying along working time

of WSN. According to the test, N4 has the lowest lifetime (1.95 hours), which is

4.8% higher than the estimated result. N5 is in charge of forwarding packets only

157


Table

6.6:Neig

hborhoodtable

andRSScompariso

nsbetw

eenrea

lmeasurem

entandsim

ulatio

nresu

lts:outdoorscen

ario

.

NodeID

NeighborID

and

Avera

geRSSs(d

Bm)

Accnb

RSS

ME,STD

1BS(-6

1.00),2

(-61.50),3

(-67.00),4

(-75.00),5

(-70.00),6

(-79.00),7

(-78.00),8

(-80.00)

BS(-6

5.65),2

(-62.26),3

(-68.25),4

(-71.40),5

(-74.47),6

(-76.41),7

(-77.99),8

(-81.13)

2BS(-7

7.00),1

(-62.30),3

(-58.00),4

(-68.00),5

(-76.00),6

(-74.00),7

(-76.00),8

(-75.00),9

(-86.00)

BS(-7

2.45),1

(-62.25),3

(-62.19),4

(-67.69),5

(-75.17),6

(-74.53),7

(-76.29),8

(-78.85),9

(-88.02)

3BS(-5

7.00),1

(-68.50),2

(-59.00),4

(-62.65),5

(-66.00),6

(-71.00),7

(-75.00),8

(-73.00),9

(-86.00)

BS(-6

8.11),1

(-68.24),2

(-62.20),4

(-61.15),5

(-68.86),6

(-72.11),7

(-74.51),8

(-72.31),9

(-80.15)

4BS(-7

9.00),1

(-74.00),2

(-67.00),3

(-63.00),5

(-70.00),6

(-71.00),7

(-70.00),8

(-77.00),9

(-85.00)

BS(-7

1.18),1

(-71.44),2

(-67.69),3

(-61.17),5

(-64.31),6

(-69.24),7

(-72.45),8

(-84.86),9

(-85.83)

ME

3=

2.45

5BS(-7

3.00),1

(-75.50),2

(-73.00),3

(-67.00),4

(-68.00),6

(-63.00),7

(-69.00),9

(-83.00)

99%

BS(-7

4.42),1

(-74.57),2

(-72.19),3

(-68.87),4

(-64.31),6

(-61.96),7

(-68.21),8

(-69.44),9

(-83.56)

STD

3=

2.45

6BS(-7

7.00),1

(-81.00),2

(-74.00),3

(-69.00),4

(-70.33),5

(-65.00),7

(-60.00),8

(-64.00),9

(-78.00)

BS(-7

6.27),1

(-76.39),2

(-74.51),3

(-72.11),4

(-69.24),5

(-61.99),7

(-62.44),8

(-66.05),9

(-79.89)

7BS(-7

9.00),1

(-76.00),2

(-72.00),3

(-76.00),4

(-67.00),5

(-69.00),6

(-59.00),8

(-63.00),9

(-63.00)

BS(-7

7.81),1

(-77.96),2

(-76.42),3

(-74.55),4

(-72.46),5

(-68.21),6

(-62.45),8

(-62.08),9

(-68.10)

81(-8

0.00),2

(-76.00),3

(-78.00),4

(-86.00),6

(-65.00),7

(-61.00),9

(-61.00)

1(-8

0.95),2

(-88.60),3

(-73.21),4

(-84.86),5

(-69.44),6

(-65.95),7

(-62.09),9

(-62.07)

92(-8

5.00),3

(-79.00),4

(-85.00),5

(-82.00),6

(-77.00),7

(-67.00),8

(-61.00)

2(-8

8.21),3

(-80.23),4

(-85.83),5

(-83.56),6

(-69.72),7

(-68.09),8

(-62.08)

158


for N6, therefore within the same time it consumes less energy than N4. While N2

123457

8

9

6

123457

8

9

6(a) iMOST solution with the lowest cost

(b) Real topology

Figure 6.12: Topology comparison: (a) One of the eight solutions generated byiMOST for outdoor environment. (b) The topology of real deployment.

159


0 10 20 30 40 50 60 70 80-100

-90

-80

-70

-60

-50

-40

-30

Measured value index aligned in the table

RSS

(dB

m)

Real measurementSimulation result depth=3

Figure 6.13: RSS comparison between real measurement and simulation result withdepth = 3.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

10

20

30

40

50

60

70

80

90

100

Working time (hours)

Rem

aini

ng e

nerg

y (%

)

node 4node 5node 2

Figure 6.14: Remaining energy of N4, N5 and N2 along the working time of WSN.

sends packets only for itself, compared with N4 and N5, it has the longest battery

life.

160

6.5. Conclusion

1 2 3 4 5 6 7 8 926

28

30

32

34

36

38

Node ID

Num

ber o

f pac

kets

Arrived packetsSent packets

Figure 6.15: Comparing the number of arrived packet and the number of sent packetfor each node in the deployment.

Fig. 6.15 compares packet delivery status in a period of 0.5 hour for the topology.

DPd= 99% which is better than the estimated result of 98%. N9 and N8, which

perform similarly as N7 in the indoor measurement, have the highest packet loss

rate (10%), because they also need more number of hops to reach BS compared

with other nodes. The average packet latency is around 2.6 s, therefore Dpl ≈ 0.97

which is slightly less (2%) than the estimated result.

6.5 Conclusion

The whole methodology introduced through this thesis is comprehensively validated

in this chapter, by launching real WSN deployments.

161


3D environment models are constructed automatically for both indoor and

outdoor scenarios. Real deployments for ambient monitoring are realized for target

regions. Experiments are carefully designed by programming the sensor nodes coping

with the application requirements, and the placements of nodes strictly follow the

planned topologies generated by iMOST.

The experimental data are categorized and compared in terms of environment

reconstruction accuracy, routing table accuracy, neighborhood table accuracy, RSS

accuracy, lifetime and packet delivery status. Both applications show the potentials

of the proposed algorithm and the developed planning tool to fulfill user requirements

with optimized and practical performance.

162

Chapter 7

Conclusions and future works

7.1 Conclusions

In this thesis, a series of novel methodologies have been proposed, in terms of

environment modelling, metric modelling, network evaluation and multi-objective

optimized planning, for efficiently planning reliable WSNs. A planning tool iMOST

is developed by integrating those proposed methodologies to assist WSN designers

conveniently constructing and evaluating network topology for outdoor and indoor

environments.

The proposed automatic 3D outdoor and indoor environment modelling

methodology liberates WSN designers and wireless communication engineers from

the traditional time consuming and costly approaches, in which environment is

usually purchased from professional GIS companies or manually reconstructing

from field measurements. With the best average accuracy of 76.1% for outdoor

image understanding within one hour and 97% accuracy for indoor recognition in

less than one minute, this method is capable to recognize and segment objects

from images pixel wisely with sufficient accuracy. The 3D vectorization procedure

eliminates redundant information thus objects are stored succinctly to reduce

memory occupations. Moreover we also prove its flexibility to be applied for

different outdoor environments by training customized environment databases and

the robustness for indoor reconstruction. To our best knowledge, this is the first

time that image understanding algorithm being applied to automatically reconstruct

environment database for signal propagation and network planning purpose. This

novel approach allows reconstructing large scale 3D map in a very short time,

accurately and for free.

Based on the practical and accurate 3D modelling, the ray tracing engine is

developed to tracing the radio and sensing signal path. The experimental results

Chapter 7. Conclusions and future works

proved the improvement on computation efficiency (in average 335%) by using

the kd-tree space division algorithm and modified polar sweep algorithm. The

radio propagation model is proposed for ray tracing engine, which emphasizes not

only the materials of obstacles but also their locations along the signal path and

the performance is evaluated through comparison with both indoor and outdoor

measurement data, and the mean error is less than 2.2 dB in outdoor test and

less than 2.42 dB for indoor scenarios, which outperforms than the compared

state-of-the-art works. The sensing signal of sensor nodes, which are sensitive to

the obstacles, is benefit from the ray-tracing algorithm via obstacles detection.

The performance of this modelling method is robust and accurate compared with

conventional methods and experimental results imply that this methodology is

suitable for both outdoor urban scenes and indoor environments, moreover it can be

applied to GSM communication and ZigBee protocol by varying frequency parameter

in the radio propagation model.

An automatic 3D multi-objective optimization WSN planning algorithm is

proposed in this work. More comprehensive metrics (connectivity, coverage, cost,

lifetime, packet latency and packet drop rate) are modeled practically compared

with other works, especially 3D ray tracing method are used to model the radio

link and sensing signal which are sensitive to the obstruction of obstacles; routing

of network is constructed by using AODV protocol; the network longevity, packet

delay and packet drop rate are obtained via simulating practical events in WSNet

simulator, which to the best of our knowledge, is the first time that network simulator

is involved in a planning algorithm. Moreover the multi-objective optimization

methodology is developed to cater the characteristics of WSNs. The individual

length is changeable so that the cost can be optimized, meanwhile crossovers

and mutations are designed to eliminate invalid modifications to improve the

computation efficiency. The capability of providing multiple optimized solutions

simultaneously allows users making their own decisions, and the results are more

comprehensive optimized compared with other state-of-the-art algorithms.

iMOST is developed by integrating the introduced novel algorithms, to assist

WSN designers efficiently planning reliable WSNs for different configurations.

iMOST features with convenient operation with user-friend vision system allow users

configuring the network properties freely; It supports the efficient and automatic

164

7.2. Future works

3D database reconstruction algorithm and fast 3D objects design for both indoor

and outdoor environments; The multiple multi-objective optimized 3D deployment

solutions in the 3D space and the corresponding evaluated performance are visually

presented to users; and the NPM of iMOST is available online as well as the source

codes of the other two rebuilt advanced heuristics. Therefore WSN designers will be

benefit from this tool on efficiently constructing environment database, practically

and efficiently planning reliable WSNs for both outdoor and indoor applications,

efficiently and accurately estimate the performance of a WSN. With the open source

codes, they are also able to compare their algorithms with ours to make contributions

to this academic field.

The whole methodology introduced through this thesis is comprehensively

validated, by launching real WSN deployments for both indoor and outdoor

environment. Experiments are carefully designed by programming the sensor nodes

coping with the application requirements, and the placements of nodes strictly

follow the planned topologies generated by iMOST. The environment reconstruction

accuracy, routing table accuracy, neighborhood table accuracy, RSS accuracy,

lifetime and packet delivery status are computed and analyzed through comparisons.

The results indicate that the proposed methodologies and the developed planning

tool iMOST are able to assist WSN designers efficiently planning reliable and

optimized WSN topology for both indoor and outdoor scenarios.

7.2 Future works

In the future, this work can be continued from different aspects: First of all, the

image understanding result for indoor environment can be improved. By observing

different types of line segments with different thickness, the materials of internal

walls can be distinguished and different dielectric parameters can automatically

assigned instead of using uniform parameter by the current work. More training

database should be constructed for different kinds of environments. Besides the

choices of some parameters in outdoor image understanding are arbitrary, which

leads to uncertainty in the performance and hence deep insight study over those

parameters will contribute to the work.

165


Secondly, computer science knowledge shall be used in the future to optimize

the computation method for large scale network planning and large scenario

demonstration. When scenario becomes larger, more points are involved in the

computation which will increase time spent on searching proper placements. One

proposal to this problem is the partition computation: each detected region is

extracted and only the areas with high deployment possibility are concerned. By

doing so, the planning algorithm does not have to traverse the entire environment

database and the speed is improved.

There are several issues that we have not yet explored in this work, including

planning mobile sensor networks, the study of human movements on the network

performance and the interference from other wireless signals over the deployed

environment and the antenna radiation pattern towards the radio propagation

modelling. Therefore more efforts should be done to make contribution to this

academic field and bring more benefits to WSN designers.

7.3 Publications based on this work

This section shows the complete list of publications resulting from this thesis:

Refereed journal papers

1. Danping He, Guixuan Liang, Jorge Portilla, Teresa Riesgo, A Novel Method

for Radio Propagation Simulation Based on Automatic 3D Environment

Reconstruction. Radioengineering. 21 - 1, pp. 985 - 992. 12/2012. ISSN

1210-2512. (Invited paper)

2. Danping He, Gabriel Mujica, Guixuan Liang, Jorge Portilla, Teresa Riesgo,

Radio Propagation Modeling and Real Test of ZigBee Based Indoor Wireless

Sensor Networks. Submitted to JSA. (Invited paper by Journal of

Systems Architecture: Embedded Software Design (JSA))

3. Danping He, Gabriel Mujica, Jorge Portilla, Teresa Riesgo, Modelling

and planning reliable wireless sensor networks based on multi-objective

optimization genetic algorithm with changeable length. Submitted to Journal

of Heuristics.

166

7.3. Publications based on this work

Refereed conference papers

1. Danping He, Nathalie Mitton, David Simplot-Ryl, An Energy Efficient

Adaptive HELLO Algorithm for Mobile Ad Hoc Networks. In Proceedings

of the 16th ACM international conference on Modeling, analysis & simulation

of wireless and mobile systems (MSWiM ’13), pp. 65 - 72, 2013.

2. Danping He, Gabriel Mujica, Guixuan Liang, Jorge Portilla, Teresa Riesgo,

Radio Propagation Modeling and Measurements for ZigBee Based Indoor

Wireless Sensor Networks. In Proceedings of the Jornadas de Computaci�n

empotrada, pp. 98 - 103, 2013. ��Best paper award, invited to publish

on JSA��

3. Danping He, Jorge Portilla, Teresa Riesgo, A 3D Multi-objective

Optimization Planning Algorithm for Wireless Sensor Networks. Published

in IECON, 2013.

4. Danping He, Guixuan Liang, Jorge Portilla, Teresa Riesgo, A Novel Method

for Radio Propagation Simulation Based on Automatic 3D Environment

Reconstruction. In Proc. 6th European Conf. Antennas and Propagation

(EUCAP), pp. 1445 - 1449, 2012. ��Invited to publish at special issue

of Radioengineering Journal��

5. Danping He, Gabriel Mujica, Jorge Portilla, Teresa Riesgo, Simulation Tool

and Case Study for Planning Wireless Sensor Network. In Proceedings of

Annual Conference of the IEEE Industrial Electronics Society (IECON), pp.

6028 - 6032, 2012.

6. Guixuan Liang, Danping He, Jorge Portilla, Teresa Riesgo, A Hardware In

The Loop Design Methodology For FPGA System and Its Application To

Complex Functions. In Proceedings of VLSI, Design, Automation and Test

(VLSI-DAT ), pp. 1 - 6, 2012.

7. Guixuan Liang, Danping He, Jorge Portilla, Teresa Riesgo, Functional

Validation of MB-OFDM System Using HW-in the loop. In Proceedings of

Conference on Design of Circuits and Integrated Systems (DCIS), pp. 131 -

136, 2012

167


8. Guixuan Liang, Danping He, Eduardo de la Torre, Teresa Riesgo,

Low-power, High-speed FFT Processor for MB-OFDMUWB Application.

Microtechnologies for the New Millennium 2011 (SPIE), 2011.

7.4 Implementation of this work

This work is implemented in the Working Package 5 of European project

WSN-DPCM, which is funded by the ARTEMIS Joint Undertaking (the European

technology platform representing the field of advanced research and technology for

embedded intelligence and systems), national authorities and European partner

companies with a total amount of 3.4 million euros. It is officially launched in

October 2011 and lasts for 36 months [142]. With the cooperation of several

technical universities and companies from Spain, Italy, Lithuania and Greece,

the project targets to address the WSN deployment, testing, and maintenance

challenging issues by developing an integrated platform for smart environments that

will comprise a middleware for heterogeneous wireless technologies as well as an

integrated engineering tool for quick system development, a planning tool and a

commissioning & maintenance tool for expert and non-expert users. The Working

Package 5 (WP5) of the WSN-DPCM project aims at developing the WSN visual

modeling, simulation and deployment tool for the optimal target WSN deployment

scheme definition. In the proposal, the planning tool should have rich graphical

end-user interface with high level of automation, supporting various proposals such

as models of WSN motes, network topology, communication models, environmental

conditions and deployment environments. With the assistance of the planning

tool, WSN developers are expected to be facilitated during the WSN distributed

application design for investigating alternative deployment schemes in a transparent

and guided way, thus saving efforts and time.

168

Bibliography

[1] D. Estrin, R. Govindan, J. Heidemann, and S. Kumar. Next century

challenges: scalable coordination in sensor networks. In Proceedings of the

5th annual ACM/IEEE international conference on Mobile computing and

networking, MobiCom ’99, pages 263–270, New York, NY, USA, 1999. ACM.

(Cited on page 1.)

[2] Zigbee Alliance. Zigbee protocol. http://www.zigbee.org/. (Cited on

page 1.)

[3] Bluetooth. https://www.bluetooth.org/. (Cited on page 1.)

[4] UWB. Iso/iec 26907:2009 information technology – telecommunications and

information exchange between systems – high-rate ultra-wideband phy and

mac standard, 2009. (Cited on page 1.)

[5] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless

sensor networks: a survey. Computer Networks, 38(4):393–422, 2002. (Cited

on page 1.)

[6] J. Yick, B. Mukherjee, and D. Ghosal. Wireless sensor network survey.

Computer Networks, 52(12):2292 – 2330, 2008. (Cited on page 1.)

[7] C. Buratti, A. Conti, D. Dardari, and R. Verdone. An overview on wireless

sensor networks technology and evolution. Sensors, 9(9):6869–6896, 2009.

(Cited on page 1.)

[8] P. Naz, S. Hengy, and P. Hamery. Soldier detection using unattended acoustic

and seismic sensors. In Proceedings of SPIE, Baltimore, Maryland, USA, 2012.

(Cited on page 3.)

[9] J. Chang, W. Mendyk, L. Thier, P. Yun, A. LaRow, S. Shaw, and

W. Schoenborn. Early attack reaction sensor (EARS), a man-wearable gunshot

detection system. In Proceedings of Society of Photo-Optical Instrumentation

Engineers (SPIE) Conference Series, volume 6201 of Society of Photo-Optical

Instrumentation Engineers (SPIE) Conference Series, Orlando (Kissimmee),

FL, USA, June 2006. (Cited on page 3.)

[10] H. E. de Bree and J. W. Wind. The acoustic vector sensor: a versatile

battlefield acoustics sensor. In Proceedings of the Society of Photo-

Optical Instrumentation Engineers (SPIE) Conference Series, volume 8047

of Society of Photo-Optical Instrumentation Engineers (SPIE) Conference

Series, Orlando, Florida, USA, May 2011. (Cited on page 3.)

Bibliography

[11] G. de Mel, T. Pham, P. Sullivan, K. Grueneberg, W. Vasconcelos, and

T. Norman. Intent-based resource deployment in wireless sensor networks. In

Proceedings of the Society of Photo-Optical Instrumentation Engineers (SPIE)

Conference Series, volume 8389 of Society of Photo-Optical Instrumentation

Engineers (SPIE) Conference Series, Baltimore, Maryland, USA, May 2012.

(Cited on page 3.)

[12] R. T. Zehr, S. K. Holland, and G. Laufer. A low-cost remote chemical sensor

for e-uav platforms. In Proceedings of SPIE Defense & Security Symposium,

Orlando, Florida USA, 2007. (Cited on page 4.)

[13] L. Bukshpun, R. D. Pradhan, V. Tun, N.and Esterkin, and G. Tomczyk.

Novel optical sensor system for missile canisters continuous monitoring. In

Proceedings of SPIE, 2007. (Cited on page 4.)

[14] E. Goldoni and P. Gamba. W-tremors, a wireless monitoring system for

earthquake engineering. In Proceedings of IEEE Workshop on Environmental

Energy and Structural Monitoring Systems (EESMS 2010), pages 26–31,

Taranto, Italy, 2010. (Cited on page 4.)

[15] J. Wong, J. Goethals, and B. Stojadinovic. Wireless sensor seismic response

monitoring system implemented on top of neesgrid. In Proceedings of SPIE

5768, Health Monitoring and Smart Nondestructive Evaluation of Structural

and Biological Systems IV, 74, volume 5768, pages 74–84, San Diego, CA,

USA, 2005. (Cited on pages 4 and 5.)

[16] P. Zhang, C. M. Sadler, S. A. Lyon, and M. Martonosi. Hardware design

experiences in zebranet. In Proceedings of the 2nd international conference on

Embedded networked sensor systems, SenSys ’04, pages 227–238, New York,

NY, USA, 2004. ACM. (Cited on page 5.)

[17] G. Werner-Allen, K. Lorincz, M. Welsh, O. Marcillo, J. Johnson, M. Ruiz,

and J. Lees. Deploying a wireless sensor network on an active volcano. IEEE

Internet Computing, 10(2):18–25, 2006. (Cited on pages 5, 6 and 105.)

[18] Tmote Sky, 2009. (Cited on page 6.)

[19] G. Tolle, J. Polastre, R. Szewczyk, D. Culler, N. Turner, K. Tu, S. Burgess,

T. Dawson, P. Buonadonna, D. Gay, and W. Hong. A macroscope in the

redwoods. In Proceedings of the 3rd international conference on Embedded

networked sensor systems, SenSys ’05, pages 51–63, New York, NY, USA,

2005. ACM. (Cited on page 6.)

[20] A. Chandrakasan, R. Amirtharajah, S. Cho, J. Goodman, G. Konduri,

J. Kulik, W. Rabiner, and A. Wang. Design considerations for distributed

170

Bibliography

microsensor systems. In Proceedings of the IEEE Custom Integrated Circuits,

pages 279–286, San Diego, CA, USA, 1999. (Cited on page 6.)

[21] Forest fire detection system. http://www.libelium.com/wireless_sensor_networks_to_detec_forest_fires/. (Cited on page 6.)

[22] A. Cerpa, J. Elson, D. Estrin, L. Girod, M. Hamilton, and J. Zhao. Habitat

monitoring: application driver for wireless communications technology.

SIGCOMM Comput. Commun. Rev., 31(2 supplement):20–41, April 2001.

(Cited on page 6.)

[23] Alert system website. http://www.alertsystems.org/. (Cited on

page 7.)

[24] I. Vasilescu, K. Kotay, D. Rus, M. Dunbabin, and P. Corke. Data collection,

storage, and retrieval with an underwater sensor network. In Proceedings of the

3rd international conference on Embedded networked sensor systems, SenSys

’05, pages 154–165, New York, NY, USA, 2005. ACM. (Cited on page 7.)

[25] J. McCulloch, P. McCarthy, S. M. Guru, W. Peng, D. Hugo, and A. Terhorst.

Wireless sensor network deployment for water use efficiency in irrigation.

In Proceedings of the workshop on Real-world wireless sensor networks,

REALWSN ’08, pages 46–50, New York, NY, USA, 2008. ACM. (Cited on

page 7.)

[26] G. Sklivanitis, J. Kimionis, and E. Kampianakis. Towards precision

agriculture: Building a soil wetness multi-hop wsn from first principles. In

Proceedings of the Second International Workshop in Sensing Technologies

in Architecture, Forestry and Environment (ECOSENSE), Belgrade, Serbia,

2011. (Cited on page 7.)

[27] J. Xia, Z. Tang, X. Shi, L. Fan, and H. Li. An environment monitoring system

for precise agriculture based on wireless sensor networks. In Proceedings of

the Seventh International Conference on Mobile Ad-hoc and Sensor Networks

(MSN 2011), pages 28–35, Beijing, China, 2011. (Cited on page 7.)

[28] H. Alemdar and C. Ersoy. Wireless sensor networks for healthcare: A survey.

Computer Networks, 54(15):2688 – 2710, 2010. (Cited on page 7.)

[29] T. Hori and Y. Nishida. Ultrasonic sensors for the elderly and caregivers

in a nursing home. In Proceedings of the 7th International Conference on

Enterprise Information Systems (ICEIS 2005), pages 110–115, Miami, USA,

2005. (Cited on pages 7 and 8.)

171

Bibliography

[30] Z. Pang, Q. Chen, and L. Zheng. A pervasive and preventive healthcare

solution for medication noncompliance and daily monitoring. In Proceedings

of the 2nd International Symposium on Applied Sciences in Biomedical and

Communication Technologies (ISABEL 2009), pages 1–6, Bratislava, Slovakia,


[31] S. Yoo, P. K. Chong, T. Kim, J. Kang, D. Kim, C. Shin, K. Sung, and

B. Jang. Pgs: Parking guidance system based on wireless sensor network.

In Proceedings of the 3rd International Symposium on Wireless Pervasive

Computing (ISWPC 2008), pages 218–222, Santorini, Greece, 2008. (Cited

on page 9.)

[32] Smartsantander. http://www.smartsantander.eu/. (Cited on pages 9,

14 and 106.)

[33] J. Rousselot, Ph. Dallemagne, and J.-D. Decotignie. Deployments of wireless

sensor networks performed by csem. In Proceedings of COGnitive systems with

Interactive Sensors (COGIS 2009), Paris, France, 2009. (Cited on page 12.)

[34] Ns-2. http://www.isi.edu/nsnam/ns/. (Cited on page 12.)

[35] Omnet++. http://www.omnetpp.org/. (Cited on page 12.)

[36] A. Fraboulet, G. Chelius, and E. Fleury. Worldsens: development and

prototyping tools for application specific wireless sensors networks. In

Proceedings of the 6th international conference on Information processing in

sensor networks, IPSN ’07, pages 176–185, New York, NY, USA, 2007. ACM.

(Cited on page 12.)

[37] P. Levis, N. Lee, M. Welsh, and D. E. Culler. Tossim: accurate and scalable

simulation of entire tinyos applications. In Proceedings of SenSys, pages

126–137, Los Angeles, California, USA, 2003. (Cited on page 12.)

[38] M. Korkalainen and M. Sallinen. A survey of rf-propagation simulation

tools for wireless sensor networks. In Proceedings of the Fourth Int

Sensor Technologies and Applications (SENSORCOMM) Conf, pages 342–347,

Venice, Italy, 2010. (Cited on pages 13 and 74.)

[39] Edxpro. http://www.edx.com/products/signalpro.html. (Cited on

page 13.)

[40] Winprop: Software tool for the planning of radio communication networks

(terrain, urban, indoor & tunnel). http://www.awe-communications.com/. (Cited on page 13.)

172

Bibliography

[41] Cindoor. http://www.gsr.unican.es/cindoor/. (Cited on page 13.)

[42] M. T. Kouakou, S. Yamamoto, K. Yasumoto, and M. Ito. Cost-efficient

deployment for full-coverage and connectivity in indoor 3d wsns. In Proceedings

of IPSJ Dicomo 2010, 2010. (Cited on pages 13, 20, 101, 124 and 130.)

[43] D. Jourdan and O.L. de Weck. Layout optimization for a wireless sensor

network using a multi-objective genetic algorithm. In Proceedings of IEEE 59th

Vehicular Technology Conference (VTC 2004), volume 5, pages 2466–2470

Vol.5, Los Angeles, CA, USA, 2004. (Cited on pages 14, 20 and 130.)

[44] Citysense website. https://www.sensenetworks.com/products/macrosense-technology-platform/citysense/. (Cited on pages 14

and 106.)

[45] M. Ahlberg, V. Vlassov, and T. Yasui. Router placement in wireless sensor

networks. In Proceedings of the IEEE International Conference on Mobile

Adhoc and Sensor Systems (MASS 2006), pages 538–541, Vancouver, BC,

Canada, 2006. (Cited on pages 14 and 106.)

[46] S. Misra, S. D. Hong, G. Xue, and J. Tang. Constrained relay node placement

in wireless sensor networks: Formulation and approximations. Networking,

IEEE/ACM Transactions on, 18(2):434–447, 2010. (Cited on pages 15 and 20.)

[47] S.M.S. Shams, A.H. Chowdhury, Ki-Hyung Kim, and Noh Bok Lee. A fast

approximation algorithm for relay node placement in double-tiered wireless

sensor network. In Proceedings of IEEE Military Communications Conference

(Milcom 2008), pages 1–6, San Diego, CA, USA, 2008. (Cited on pages 15

and 20.)

[48] S. Lee and M. Lee. Qrmsc: Efficient qos-aware relay node placement in wireless

sensor networks using minimum steiner tree on the convex hull. In Proceedings

of The International Conference on Information Networking (ICOIN 2013),

pages 36–41, Bangkok, Thailand, 2013. (Cited on pages 15 and 20.)

[49] S. Kim, J. Ko, J. Yoon, and H. Lee. Multiple-objective metric for placing

multiple base stations in wireless sensor networks. In Proceedings of the 2nd

International Symposium on Wireless Pervasive Computing (ISWPC ’07),

pages –, San Juan, Puerto Rico, 2007. (Cited on pages 15 and 20.)

[50] Y. Huang, P. Hsiu, W. Chu, K. Hung, A. Pang, T. Kuo, M. Di, and H. Fang.

An integrated deployment tool for zigbee-based wireless sensor networks.

In Proceedings of IEEE/IFIP International Conference on Embedded and

Ubiquitous Computing (EUC ’08), volume 1, pages 309–315, Shanghai, China,

2008. (Cited on pages 16, 17 and 20.)

173

Bibliography

[51] A. Guinard, M.S. Aslam, D. Pusceddu, S. Rea, A. McGibney, and D. Pesch.

Design and deployment tool for in-building wireless sensor networks: A

performance discussion. In Proceedings of IEEE 36th Conference on Local

Computer Networks (LCN 2011), pages 649–656, Bonn, Germany, 2011. (Cited

on pages 16, 18, 20 and 101.)

[52] A. McGibney, A. Guinard, and D. Pesch. Wi-design: A modelling and

optimization tool for wireless embedded systems in buildings. In Proceedings

of IEEE 36th Conference on Local Computer Networks (LCN 2011), pages

640–648, Bonn, Germany, 2011. (Cited on pages 16, 20 and 101.)

[53] V. Akbarzadeh, C. Gagne, M. Parizeau, M. Argany, and M.A. Mostafavi.

Probabilistic sensing model for sensor placement optimization based on

line-of-sight coverage. Instrumentation and Measurement, IEEE Transactions

on, 62(2):293–303, 2013. (Cited on pages 19 and 20.)

[54] S. Xiong, L. Yu, H. Shen, C. Wang, and W. Lu. Efficient algorithms for

sensor deployment and routing in sensor networks for network-structured

environment monitoring. In Proceedings of the IEEE INFOCOM 2012, pages

1008–1016, Orlando, FL, USA, 2012. (Cited on pages 19 and 20.)

[55] L. Liu and H. Ma. On coverage of wireless sensor networks for rolling terrains.

Parallel and Distributed Systems, IEEE Transactions on, 23(1):118–125, 2012.

(Cited on pages 19 and 20.)

[56] S. Qian, P. Guo, and T. Jiang. A novel lifetime-enhanced deployment strategy

for chain-type wireless sensor networks. In Proceedings of IEEE International

Conference on Communications (ICC 2012), pages 513–517, Ottawa, ON,

Canada, 2012. (Cited on pages 19 and 20.)

[57] M.Z.A. Bhuiyan, J. Cao, and G. Wang. Deploying wireless sensor networks

with fault tolerance for structural health monitoring. In Proceedings of IEEE

8th International Conference on Distributed Computing in Sensor Systems

(DCOSS 2012), pages 194–202, Hangzhou, China, 2012. (Cited on pages 19

and 20.)

[58] Y. Chen, C. Chuah, and Q. Zhao. Sensor placement for maximizing

lifetime per unit cost in wireless sensor networks. In Proceedings of IEEE

Military Communications Conference (MILCOM 2005), pages 1097–1102 Vol.

2, Atlantic city, New jersey, USA, 2005. (Cited on page 20.)

[59] A. Efrat, Sariel H., and J. S. B. Mitchell. Approximation algorithms

for two optimal location problems in sensor networks. In Proceedings of

2nd International Conference on Broadband Networks (BROADNETS 2005),

Boston, Massachusetts, USA, 2004. (Cited on pages 19 and 101.)

174

Bibliography

[60] X. Cheng, D. Du, L. Wang, and B. Xu. Relay sensor placement in wireless

sensor networks. Wirel. Netw., 14(3):347–355, June 2008. (Cited on pages 19

and 101.)

[61] S. Poduri, S. Pattem, B. Krishnamachari, and G. S. Sukhatme. Sensor network

configuration and the curse of dimensionality. In Proceedings of the Third

IEEE Workshop on Embedded Networked Sensors, 2006. (Cited on pages 19

and 101.)

[62] H. Mayer. Automatic object extraction from aerial imagery - a survey focusing

on buildings. Computer Vision and Image Understanding, 74(2):138 – 149,


[63] C. Heipke, H. Mayer, C. Wiedemann, and O. Jamet. Evaluation of automatic

road extraction. In Proceedings of International Archives of Photogrammetry

and Remote Sensing, pages 47–56, 1997. (Cited on page 28.)

[64] M. Chikr El-Mezouar, N. Taleb, K. Kpalma, and J. Ronsin. A high-resolution

index for vegetation extraction in ikonos images. In Proceedings of the SPIE,

volume 7824, pages 78242A–78242A–9, 2010. (Cited on page 28.)

[65] N. Haala, C. Brenner, and A. Karl-heinrich. 3d urban gis from laser altimeter

and 2d map data. In Proceedings of International Archives of Photogrammetry

and Remote Sensing, pages 339–346, 1998. (Cited on page 28.)

[66] F. Rottensteiner and Ch. Briese. Automatic generation of building models

from lidar data and the integration of aerial images. In Proceedings of

ISPRS working group III/3 workshop on �3-D reconstruction from airborne

laserscanner and InSAR data�, pages 174–180, Dresden, Germany, 2003.

(Cited on page 29.)

[67] D. G. Jones and J. Malik. A computational framework for determining

stereo correspondence from a set of linear spatial filters. In G. Sandini,

editor, Computer Vision?ECCV’92, volume 588 of Lecture Notes in Computer

Science, pages 395–410. Springer Berlin Heidelberg, 1992. (Cited on pages 29

and 39.)

[68] H. Lin, J. Gao, Y. Zhou, G. Lu, M. Ye, C. Zhang, L. Liu, and R. Yang.

Semantic decomposition and reconstruction of residential scenes from lidar

data. ACM Trans. Graph., 32(4):66:1–66:10, July 2013. (Cited on page 29.)

[69] A. Laurentini. The visual hull concept for silhouette-based image

understanding. Pattern Analysis and Machine Intelligence, IEEE

Transactions on, 16(2):150–162, 1994. (Cited on page 29.)

175

Bibliography

[70] W. Matusik, C. Buehler, R. Raskar, S. J. Gortler, and L. McMillan.

Image-based visual hulls. In Proceedings of the 27th annual conference

on Computer graphics and interactive techniques, SIGGRAPH ’00, pages

369–374, New York, NY, USA, 2000. ACM Press/Addison-Wesley Publishing

Co. (Cited on page 30.)

[71] K.N. Kutulakos and S.M. Seitz. A theory of shape by space carving.

In Proceedings of the Seventh IEEE International Conference on Computer

Vision (ICCV’99), volume 1, pages 307–314 vol.1, Kerkyra, Greece, 1999.

(Cited on page 31.)

[72] M. Levoy and P. Hanrahan. Light field rendering. In Proceedings of the

23rd annual conference on Computer graphics and interactive techniques,

SIGGRAPH ’96, pages 31–42, New York, NY, USA, 1996. ACM. (Cited on

pages 30 and 31.)

[73] M. R. Oswald, E. Toppe, and D. Cremers. Fast and globally optimal single

view reconstruction of curved objects. In Proceedings of the 24th IEEE

Conference on Computer Vision and Pattern Recognition (CVPR 2012), pages

534–541, Providence, Rhode Island, USA, 2012. (Cited on page 30.)

[74] Mukta Prasad and Andrew W. Fitzgibbon. Single view reconstruction of

curved surfaces. In CVPR (2), pages 1345–1354, 2006. (Cited on page 30.)

[75] A. Saxena, M. Sun, and A.Y. Ng. 3-d reconstruction from sparse views using

monocular vision. In Proceedings of IEEE 11th International Conference on

Computer Vision (ICCV 2007), pages 1–8, Rio de Janeiro, Brazil, 2007. (Cited

on page 31.)

[76] A. Saxena, M. Sun, and A.Y. Ng. Learning 3-d scene structure from a

single still image. In Proceedings of IEEE 11th International Conference on

Computer Vision (ICCV 2007), pages 1–8, Rio de Janeiro, Brazil, 2007. (Cited

on page 31.)

[77] C. Zou, J. Liu, and J. Liu. Precise 3d reconstruction from a single image.

In Proceedings of the Asian Conference on Computer Vision (ACCV 2012),

pages 271–282, Daejeon, Korea, 2012. (Cited on page 32.)

[78] L. He and H. Liu. Shape context for image understanding. In Proceedings of the

5th WSEAS international conference on Signal, speech and image processing,

SSIP’05, pages 276–281, Stevens Point, Wisconsin, USA, 2005. World Scientific

and Engineering Academy and Society (WSEAS). (Cited on page 34.)

176

Bibliography

[79] S. Belongie, J. Malik, and J. Puzicha. Shape matching and object recognition

using shape contexts. Pattern Analysis and Machine Intelligence, IEEE


[80] S. Arivazhagan, R. N. Shebiah, S.S. Nidhyanandhan, and L. Ganesan. Fruit

recognition using color and texture features. Journal of Emerging Trends in

Computing and Information Sciences, 1(2):90–94, January 2010. Published by

Foundation of Computer Science. (Cited on page 35.)

[81] X. He, R.S. Zemel, and M.A. Carreira-Perpindn. Multiscale conditional

random fields for image labeling. In Proceedings of the IEEE Computer

Society Conference on Computer Vision and Pattern Recognition (CVPR

2004), volume 2, pages II–695–II–702 Vol.2, Washington, DC, USA, 2004.

(Cited on page 35.)

[82] J. Shotton, J. Winn, C. Rother, and A. Criminisi. Textonboost for image

understanding: Multi-class object recognition and segmentation by jointly

modeling texture, layout, and context. Int. J. Comput. Vision, 81(1):2–23,

January 2009. (Cited on pages 35, 37, 44, 47 and 49.)

[83] L. Breiman. Special invited paper. additive logistic regression: A statistical

view of boosting: Discussion. The Annals of Statistics, 28(2):pp. 374–377,


[84] C. Elkan. Using the Triangle Inequality to Accelerate k-Means. In Proceedings

of the 30th International Conference on Machine Learning (ICML 2013),

pages 147–153, Atlanta, GA, USA, 2003. (Cited on page 40.)

[85] J. Friedman, T. Hastie, and R. Tibshirani. Additive logistic regression: a

statistical view of boosting. Annals of Statistics, 28:2000, 1998. (Cited on

page 43.)

[86] Y. Boykov, O. Veksler, and R. Zabih. Fast approximate energy minimization

via graph cuts. IEEE Trans. Pattern Anal. Mach. Intell., 23(11):1222–1239,

November 2001. (Cited on page 44.)

[87] Y. Boykov and V. Kolmogorov. An experimental comparison of min-cut/max-

flow algorithms for energy minimization in vision. Pattern Analysis and

Machine Intelligence, IEEE Transactions on, 26(9):1124–1137, 2004. (Cited

on page 44.)

[88] P. Soille. Morphological Image Analysis: Principles and Applications.

Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2 edition, 2003. (Cited

on page 51.)

177

Bibliography

[89] R. Wilson, S. C. Clippingdale, and A. H. Bhalerao. Robust estimation of

local orientations in images using a multiresolution approach. In Proceedings

of SPIE 1360, Visual Communications and Image Processing ’90: Fifth in

a Series, 1393, volume 1360, pages 1393–1403, Lausanne, Switzerland, 1990.

(Cited on page 52.)

[90] C. Harris and M. J. Stephens. A combined corner and edge detector. In

Proceedings of the fourth Alvey Vision Conference, pages 147–152, University

of Manchester, 1988. (Cited on page 66.)

[91] East Lansing. Floor plan of east lansing, mi 48823. (Cited on pages 69 and 70.)

[92] E. DAMOSSO. Final report of COST 231. Digital Mobile Radio towards

future Generation Systems. European Comission,Bruxelles, 1999. (Cited on

pages 69, 73 and 83.)

[93] L. M. Kamarudin, R. B. Ahmad, B. L. Ong, F. Malek, A. Zakaria, and

M. A. M. Arif. Review and modeling of vegetation propagation model for

wireless sensor networks using omnet++. In Proceedings of the Second Int

Network Applications Protocols and Services (NETAPPS) Conf, pages 78–83,

Kedah, 2010. (Cited on page 73.)

[94] J. Kenyeres, S. Sajban, P. Farkas, and M. Rakus. Indoor experiment with

wsn application. In Proceedings of the 33rd Int MIPRO Convention, pages

863–866, Opatija, Croatia, 2010. (Cited on page 73.)

[95] R. M. Pellegrini, S. Persia, D. Volponi, and G. Marcone. Rf propagation

analysis for zigbee sensor network using rssi measurements. In Proceedings

of the 2nd Int Wireless Communication, Vehicular Technology, Information

Theory and Aerospace & Electronic Systems Technology (Wireless VITAE)

Conf, pages 1–5, Chennai, India, 2011. (Cited on page 74.)

[96] V. Kolar, S. Razak, P. Mahonen, and N. B. Abu-Ghazaleh. Measurement

and analysis of link quality in wireless networks: An application perspective.

In Proceedings of IEEE Conf. Computer Communications Workshops

(INFOCOM 2010), pages 1–6, San Diego, CA, USA, 2010. (Cited on page 74.)

[97] T. Chrysikos, G. Georgopoulos, and S. Kotsopoulos. Wireless channel

characterization for a home indoor propagation topology at 2.4 ghz. In

Proceedings of Wireless Telecommunications Symp (WTS 2011), pages 1–10,

New York City, NY, USA, 2011. (Cited on page 74.)

[98] K. R. Schaubach, N. J. Davis, and T. S. Rappaport. A ray tracing method

for predicting path loss and delay spread in microcellular environments. In

178

Bibliography

Proceedings of IEEE 42nd Vehicular Technology Conference, pages 932–935,

Denver, CO, USA, 1992. (Cited on page 74.)

[99] L. Schmitz, A.and Kobbelt. Wave propagation using the photon path map.

In Proceedings of International Symposium on Performance Evaluation of

Wireless Ad Hoc, Sensor, and Ubiquitous Networks (PE-WASUN�06), pages

158–161, Torremolinos, Malaga, Spain, 2006. (Cited on page 74.)

[100] S. Kim, Jr. Guarino, B. J., III Willis, T. M., V. Erceg, S. J. Fortune, R. A.

Valenzuela, L. W. Thomas, J. Ling, and J. D. Moore. Radio propagation

measurements and prediction using three-dimensional ray tracing in urban

environments at 908 mhz and 1.9 ghz. 48(3):931–946, 1999. (Cited on page 74.)

[101] A. Schmitz, T. Rick, T. Karolski, L. Kobbelt, and T. Kuhlen. Beam tracing

for multipath propagation in urban environments. In Proceedings of the 3rd

European Conf. Antennas and Propagation (EuCAP 2009), pages 2631–2635,

Berlin, Germany, 2009. (Cited on page 74.)

[102] R. Wahl and G. Wolfle. Combined urban and indoor network planning using

the dominant path propagation model. In Proceedings of the First European

Conf. Antennas and Propagation (EuCAP 2006), pages 1–6, Nice, France,


[103] H. Son and N. Myung. A deterministic ray tube method for microcellular wave

propagation prediction model. 47(8):1344–1350, 1999. (Cited on page 74.)

[104] Z. Lai, H. Song, P. Wang, H. Mu, L. Wu, and J. Zhang. Implementation

and validation of a 2.5d intelligent ray launching algorithm for large urban

scenarios. In Proceedings of the 6th European Conf. Antennas and Propagation

(EUCAP 2012), pages 2396–2400, Prague, Czech Republic, 2012. (Cited on

page 74.)

[105] T. Rautiainen, R. Hoppe, and G. Wolfle. Measurements and 3d ray tracing

propagation predictions of channel characteristics in indoor environments.

In Proceedings of the IEEE 18th International Symposium on Personal,

Indoor and Mobile Radio Communications (PIMRC 2007), pages 1–5, Athens,

Greece, 2007. (Cited on page 74.)

[106] V. Havran. Heuristic Ray Shooting Algorithms. Ph.d. thesis, Department of

Computer Science and Engineering, Faculty of Electrical Engineering, Czech

Technical University in Prague, November 2000. (Cited on page 75.)

[107] I. Wald. Realtime Ray Tracing and Interactive Global Illumination. PhD

thesis, Computer Graphics Group, Saarland University, 2004. (Cited on

page 75.)

179

Bibliography

[108] J. Bikker. Ray tracing in real -Time games. PhD thesis, NHTV University,

2012. (Cited on pages 76 and 77.)

[109] D. He, G. Liang, J. Portilla, and T. Riesgo. A novel method for radio

propagation simulation based on automatic 3d environment reconstruction.

In Proceedings of the 6th European Conf. Antennas and Propagation (EUCAP

2012), pages 1445–1449, Prague, Czech Republic, 2012. (Cited on pages 79

and 107.)

[110] D. A. McNamara, C. W. I. Pistorius, and J. A. G. Malherbe. Introduction to

the Uniform Geometrical Theory of Diffraction. Boston, MA: Artech House,


[111] Z. Yun, Z. Zhang, and M.F. Iskander. A ray-tracing method based on

the triangular grid approach and application to propagation prediction in

urban environments. Antennas and Propagation, IEEE Transactions on,

50(5):750–758, 2002. (Cited on page 87.)

[112] J. Portilla, A. de Castro, E. de la Torre, and T. Riesgo. A modular architecture

for nodes in wireless sensor networks. J. UCS, 12(3):328–339, 2006. (Cited on

pages 88 and 111.)

[113] Greenorbs. http://www.greenorbs.org/. (Cited on page 105.)

[114] H. Abelson, D. Allen, D. Coore, C. Hanson, G. Homsy, T. F. Knight Jr.,

R. Nagpal, E. Rauch, G. J. Sussman, and R. Weiss. Amorphous computing.

Commun. ACM, 43(5):74–82, May 2000. (Cited on page 106.)

[115] R Handcock, D. Swain, G. Bishop-Hurley, K. Patison, Wark T., P. Valencia,

P. Corke, and C. O’Neill. Monitoring animal behaviour and environmental

interactions using wireless sensor networks, gps collars and satellite remote

sensing. Sensors, 9(5):3586–3603, 2009. (Cited on page 106.)

[116] R. Zviedris, A. Elsts, G. Strazdins, A. Mednis, and L. Selavo. Lynxnet:

Wild animal monitoring using sensor networks. In Proceedings of the

4th International Workshop on Real-World Wireless Sensor Networks

(REALWSN 2010), volume 6511 of Lecture Notes in Computer Science, pages

170–173, Colombo, Sri Lanka, 2010. Springer. (Cited on page 106.)

[117] Xu Xu, Weifa Liang, TimWark, and Jaein Jeong. Maximizing network lifetime

via 3g gateway assignment in dual-radio sensor networks. In Proceedings of

the 37th Annual IEEE Conference on Local Computer Networks (LCN 2012),

pages 479–486, Clearwater Beach, FL, USA, 2012. (Cited on page 106.)

180

Bibliography

[118] D.l. Bimschas, S. P. Fekete, S. Fischer, H. Hellbruck, A. Kroller, R. Mietz,

M. Pagel, D. Pfisterer, K. Romer, and T. Teubler. Real-world g-lab:

Integrating wireless sensor networks with the future internet. In Proceedings of

The 6th International Conference on Testbeds and Research Infrastructures for

the Development of Networks & Communities (TridentCom 2010), volume 46,

pages 577–579, Berlin, Germany, 2010. (Cited on page 106.)

[119] C. Perkins, E. Belding-Royer, and S. Das. Ad hoc on-demand distance vector

(aodv) routing, 2003. (Cited on page 108.)

[120] D. Johnson, Y. Hu, and D. Maltz. The dynamic source routing protocol (dsr)

for mobile ad hoc networks for ipv4, 2007. (Cited on page 108.)

[121] J. Beutel, O. Kasten, and M. Ringwald. Poster abstract: Btnodes – a

distributed platform for sensor nodes. In Proceedings of the 1st International

Conference on Embedded Networked Sensor Systems, pages 292 – 293, Los

Angeles, California, USA, January 2003. ACM Press. (Cited on pages 109

and 110.)

[122] K. Dantu, M. H. Rahimi, H. Shah, S. Babel, A. Dhariwal, and G. S.

Sukhatme. Robomote: enabling mobility in sensor networks. In Proceedings

of the Fourth International Symposium on Information Processing in Sensor

Networks (IPSN 2005), pages 404–409, UCLA, Los Angeles, California, USA,


[123] M. B. McMickell, B. Goodwine, and L. A. Montestruque. Micabot: a robotic

platform for large-scale distributed robotics. In Proceedings of the 2003 IEEE

International Conference on Robotics and Automation (ICRA 2003), pages

1600–1605, Taipei, Taiwan, 2003. (Cited on page 109.)

[124] S. Bergbreiter and K. S. J. Pister. Cotsbots: an off-the-shelf platform for

distributed robotics. In Proceedings of IEEE/RSJ International Conference

on Intelligent Robots and Systems (IROS 2003)., pages 1632–1637, Nevada,

USA, 2003. (Cited on page 109.)

[125] A. Savvides and M. B. Srivastava. A distributed computation platform for

wireless embedded sensing. In Proceedings of the 20th International Conference

on Computer Design (ICCD 2002), pages 220–225, Freiburg, Germany, 2002.

(Cited on page 109.)

[126] R. M. Kling. Intel mote: An enhanced sensor network node. In Proceedings

of the International Workshop on Advanced Sensors, Structural Health

Monitoring, and Smart Structures, pages 1–6, Raiosha, Japan, November


181

Bibliography

[127] Waspmote - wireless sensor networks 802.15.4 zigbee mote. http://www.libelium.com/products/waspmote. (Cited on page 111.)

[128] Wismote. http://www.aragosystems.com/en/wisnet-item/wisnet-wismote-item.html. (Cited on page 111.)

[129] Seed-eye. http://www.evidence.eu.com/products/seed-eye.html. (Cited on page 111.)

[130] Tyndall mote. http://www.tyndall.ie/content/wireless-sensor-networks. (Cited on page 111.)

[131] B. Blaszczyszyn and B. Radunovic. Using transmit-only sensors to reduce

deployment cost of wireless sensor networks. In Proceedings of the 27th IEEE

International Conference on Computer Communications, Joint Conference of

the IEEE Computer and Communications Societies (INFOCOM 2008), pages

1202–1210, Phoenix, AZ, USA, 2008. (Cited on page 112.)

[132] M. Mafuta, M. Zennaro, A. Bagula, G. Ault, H. Gombachika, and T. Chadza.

Successful deployment of a wireless sensor network for precision agriculture

in malawi. In Proceedings of the 3rd IEEE International Conference on

Networked Embedded Systems for Every Application ( NESEA 2012 ), pages

1–7, Liverpool, UK, 2012. (Cited on page 115.)

[133] K. Langendoen, A. Baggio, and O. Visser. Murphy loves potatoes: experiences

from a pilot sensor network deployment in precision agriculture. In Proceedings

of the 20th international conference on Parallel and distributed processing,

IPDPS’06, pages 174–174, Washington, DC, USA, 2006. IEEE Computer

Society. (Cited on page 115.)

[134] S. Lau, T. Chang, S. Hu, H. Huang, L. Shyu, C. Chiu, and P. Huang. Sensor

networks for everyday use: the bl-live experience. In Proceedings of IEEE

International Conference on Sensor Networks, Ubiquitous, and Trustworthy

Computing (SUTC 2006), volume 1, pages 336–343, Taichung, Taiwan, 2006.

(Cited on page 115.)

[135] K. Romer and F. Mattern. The design space of wireless sensor networks.

Wireless Communications, IEEE, 11(6):54–61, 2004. (Cited on page 117.)

[136] F. Dai and J. Wu. On constructing k-connected k-dominating set in wireless

networks (ipdps 2005). In Proceedings of the 19th IEEE International Parallel

and Distributed Processing Symposium, pages 81a–81a, Denver, Colorado,

USA, 2005. (Cited on pages 117 and 118.)

182

Bibliography

[137] Y. Wu and Y. Li. Construction algorithms for k-connected m-dominating

sets in wireless sensor networks. In Proceedings of the 9th ACM International

Symposium on Mobile Ad Hoc Networking and Computing, MobiHoc ’08, pages

83–90, New York, NY, USA, 2008. ACM. (Cited on page 117.)

[138] X. Bai, D. Xuan, Z. Yun, T. H. Lai, and W. Jia. Complete optimal deployment

patterns for full-coverage and k-connectivity (k≤6) wireless sensor networks.

In Proceedings of the 9th ACM International Symposium on Mobile Ad Hoc

Networking and Computing, MobiHoc ’08, pages 401–410, New York, NY,

USA, 2008. ACM. (Cited on page 118.)

[139] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A fast and elitist

multiobjective genetic algorithm: Nsga-ii. Evolutionary Computation, IEEE


[140] Danping He, Jorge Portilla, and Teresa Riesgo. A 3d multi-objective

optimization planning algorithm for wireless sensor networks. In Proceedings

of the 39th Annual Conference of the IEEE Industrial Electronics Society

(IECON 2013), Vienna, Austria, 2013. (Cited on page 130.)

[141] G. Mujica, V. Rosello, J. Portilla, and T. Riesgo. Hardware-software

integration platform for a wsn testbed based on cookies nodes. In Proceedings

of the 38th Annual Conference on IEEE Industrial Electronics Society (IECON

2012), pages 6013–6018, Montreal, Quebec, Canada, 2012. (Cited on

page 143.)

[142] Wsn dpcm project. http://www.wsn-dpcm.eu/. (Cited on page 168.)

183

a novel methodology for planning reliable wireless sensor...

Documents