1

Detection and Discrimination of Sniffing and Panting Sounds of Dogs

Ophir Azulai(1), Gil Bloch(1), Yizhar Lavner (1,2), Irit Gazit (3) and Joseph Terkel (3).

1. Signals and Image Processing Lab (SIPL), Faculty of Electrical Engineering, Technion IIT, Haifa, Israel (http://www-sipl.technion.ac.il)

2. Computer Science, Tel-Hai Academic College, Upper Galilee, Israel.3. Department of Zoology, George S. Wise Faculty of Life Sciences Tel-Aviv

University, Tel Aviv 69978, Israel.

2

Abstract

Dogs’ acute sense of smell enables them to accurately identify many types of odors. This ability serves in the detection of drugs, explosives, finding people, and other special tasks. The possibility of controlling dogs’ behavior suggests many potential applications.For this purpose, an automatic system for recognition of panting and sniffing according to their corresponding acoustic signals was developed. The system consists of three main components: analysis, training and discrimination. In the analysis stage the events were detected using short-time energy and zero-crossing rate functions. Features for discrimination were extracted utilizing an autoregressive model or by means of a Mel-frequency cepstrum. A few parameters of the spectral envelope were selected, based on their discrimination ability. In the training stage, a feedforward neural network was trained, using labeled signals. Alternatively, a fuzzy K-nearest neighbor algorithm can be used. Correct identification of up to 96.8% was achieved in the discrimination stage.

3

Panting and Sniffing Signals

4

Discrimination Method 1

Identify the calibrationevents using

short time energyfunction

Train a multi-layerneural network .

Calibration Process Analyse Process

Extract AR modelfrom each event

Extract AR modelcoefficients from

each event

Extractpre-defined

parameters fromAR frequency

response

Identify the events usingshort time energy

function

Extract AR modelfrom each event

Extract AR modelcoefficients from

each event

Present theparameters to the

trained neuralnetwork

Extractpre-defined

parameters fromAR frequency

response

Use a pre-defineddecision rule

5

Events finding

Using Normalized Short Time Energy Function

6

AR model frequency response

7

Neural Network

Using three layers, feed forward network with back propagation learning method

Inpu

t V

ecto

r

Input Layer

Hidden Layer

Output Layer

8

Discrimination Method 2

Calculate theMFCC's of the

signal

Calculate theEFV's of the

signal

Classificationusing fuzzy

KNN

Decision rulefor eventlocation

Analyze Process

Identify the calibration

events

Extract MFCC'sfrom each

event

Find the mostseparable

MFCCcoefficient

Create theEFV's fromeach event

Calibration Process

9

Mel-Frequency Cepstral Coefficients

10

Event Feature Vector

11

Fuzzy KNN

Y-A

xis

X-Axis

AnalyzedFeature Vector

PantingFeatureVectors

SniffingFeatureVectors

k

j

m

j

k

j

m

jij

i

xx

xxU

xU

1

)1/(2

1

)1/(2

/1

/1

Fuzzy KNN with K=3

12

Membership FunctionClassification results using Fuzzy KNN

13

Comparison of both methods

#PhaseMethod 1Method 2

1.Event findingUsing short time energy function

Manually

2.Calculation of basic feature vector

AR model frequency response

MFCC

3.Parameters extraction

ManuallyMost Separable

4.Feature vectorExtracted parameters

EFV

5. CalibrationNeural Network Training

-

Calibration Process:

14

Comparison of both methods continued

#PhaseMethod 1Method 2

1.Event findingUsing short time energy function

-

2.Calculation of basic feature vector

AR model frequency response

MFCC

3.Feature vectorExtracted parameters

EFV

4.ClassificationNeural NetworkFuzzy KNN

5.Decision rulePre-defined threshold

Not yet defined

Analyze Process:

15

Conclusions

1. Sniffing and Panting sounds has unique signature that can discriminate between them

2. Some frequency response model can give us the desired information.

3. Event finding should combine both feature vector and some other method like energy function or ZCR.