real time image processing applied to traffic queue€¦ · web view2. signal & image...

Real Time Image Processing Applied to Traffic Queue Detection Algorithm.

1. INTRODUCTION

A major problem associated with a real-time vision-based application such as

traffic data collection is the requirement for high computational power. In traffic

applications, variation of ambient lighting, shadows, occlusion, and lane changing

movement of vehicles further complicates the task. Despite these difficulties, a vision-

based traffic data collection system has several advantages over the conventional point

based alternatives such as magnetic loops. Vision can give time-variant and spatial

information about a scene and can be recorded easily for further analysis. Ease of

installation on a present road, flexibility, and maintenance issues are among other

advantages, which encourage researchers to concentrate

To achieve real-time processing, vision-based traffic parameter extraction

techniques are based on processing key regions of the scene. The vehicles are detected in

a user-defined window by applying a background differencing technique. Since

background-based algorithms are very sensitive to ambient lighting conditions, they have

not yielded the expected results. The method proposed in this paper is based on applying

edge-detection techniques to windows placed by the user at the key regions across the

lanes.

This paper aimed to measure queue parameters accurately. The traffic queue detection algorithm has two operations:

Vehicle detection Motion detection

Previous works about queue detection using image processing are mainly focused

on measuring queue length at intersections where queues are caused by a traffic lights

system. Fatty and Sisal approach is based on computing motion and presence of vehicles

on profiles placed in strategic zones of the image. Profiles have a shape determined by

camera and geometry parameters and their output gives information about queue length.

In output of profiles is interpreted by a neural net in order to obtain better queue length

estimation. Aim of this system is to give an instantaneous precise estimation of queue

length. The vehicle detection is based on applying edge detection on these profiles.

Dept. of IT_MACE_ Venjaramoodu


2. SIGNAL & IMAGE PROCESSING APPLIED TO TRAFFIC

2.1. Digital signal

Digital systems functions more reliable, because only two states are used

(zero’s and one’s BINARY SYSTEM). The word signal is a function of independent

variables. The signal itself carries some kind of information available for observation.

By processing we mean operating in some fashion on signal to extract

some useful information. In many cases the processing will be none destructive

“TRANEFER MODE” of the given data signals.

Signal processing will be done in many ways.

1. Analog signal processing

2. Digital signal processing

3. Light(Optical)wave signal processing

4. Radio frequency signal transmission



2.1.1. Digital signal processing

Fig (2.1).Digital Signal Processing



2.2. Need for processing of traffic data

Traffic surveillance and control, traffic management, road safety and development of

transport policy.

2.3 .Traffic parameters measurable

Traffic parameters measurable are the number of active tracks(Vehicles), instantaneous velocity of a vehicle, class of a vehicle, and average velocity of a vehicle during the entire time when it was in camera’s field of view.Traffic volumes, Speed, Headways, Inter-vehicle gaps,

Vehicle classification, Origin and destination of traffic, Junction turning.



2.4. Image analysis system structure

Fig (2.2). Image analysis system structure


CCTV

camera

ADC

RAM

64k bytes

(Data bus)

Backing store

2 ¼ M-byte

16-bit

Mini-

Computer

Monitor

DAC

Printer


3. STAGES OF IMAGE ANALYSIS

Image sensors used Improved video cameras: automatic gain control, low SNR

ADC Conversion Analog video signal received from video camera is converted to

digital/binary form for processing

Pre-processing High SNR of the camera output reduces the quantity of processing

enormous data flow.

To cope with this, two methods are proposed:

1. Analyze data in real time - uneconomical

2. Stores all data and analyses off-line at low speed.

Pipeline Preprocessing does this job.

3.1 Stages in pipeline preprocessing:

(1) Spatial Averaging – contiguous pixels are averaged (convolution).

(2) Subtraction of background scene from incoming picture.

(3) Threshold – Large differences are true ‘1’, small differences are false

‘0’.

(4) Data Compression – reduces resulting data.

(5) Block buffering – collects data into blocks.

(6) Tape Interface – blocks are loaded onto a digital cassette recorder. Preprocessed picture is submitted to processor as 2-D array of numbers



Two jobs to be done

Green light on

Determine no. of vehicles moving along particular lanes and their classification by

shape and size.

Red light on

Determine the backup length along with the possibility to track its dynamics and

classify vehicles in backup



4. METHODS OF VEHICLE DETECTION

Detection of each single vehicle in the traffic scene is a very complex task due to

problems such as occlusions and shadows. In order to measure a global presence of

vehicle on the monitored road, this task can be avoided and the presence parameter can be

measured at pixel level. There are two main approaches to estimate the presence of new

objects in an image sequence taken from a static camera: background based methods and

gradient based methods. Background methods perform the difference of the current frame

with an updated background image, while the second approach is based on the detection

of vehicle edges.

Among them the good one is second strategy, which guarantees a very fast

computation and good results. In fact, considering that the road surface presents an almost

flat grey level and that vehicles generate strong discontinuities in the intensity, edges can

be a good indicator for the presence of vehicles

Many edge-detection algorithms based on gradient operators or statistical

approaches have been developed. The gradient operators are usually sensitive to noise

and are used with filtering operators to reduce the effect of noise. A statistical approach is

unable to detect thin edges, which are very likely to occur at a window of an image. The

morphological-based edge detectors have proved to be more effective than conventional

gradient-based techniques. Some commonly used morphological edge-detection

algorithms are blur-min, top-hat, and open–close operators. These operators have better

performance than gradient and facet edge operators. The authors have developed a new

morphological edge-detection operator separable morphological edge detector (SMED)

which has a lower computational requirement while having comparable performance to

other morphological operators. The reasons for adopting SMED operator in our

application are as follows.



1) SMED can detect edges at different angles, while other morphological operators are

unable to detect all kinds of edges

2) The strength of the edges detected by SMED is twice than other edge detectors. Blur-

min is unable to detect thin lines and corners, and the strength of the edges detected is

weak.

3) SMED uses separable median filtering to remove noise. Median filtering has proven to

preserve edges while removing noise. Separable median filtering has shown to have

comparable performance to the true median filtering, but requires less computational

power.

In brief, SMED, which uses compatible and easily implementable operators, has a

lower computational requirement, compared to the other morphological edge-detection

operators. Open–close has better performance than SMED operator does, but it has about

eight times more computational.



5. QUEUE DETECTION ALGORITHM

The algorithm used to detect and measure queue parameters consists of two

operations—one involving motion detection operation and the other vehicle detection

operation. These operations are applied to a profile consisting of mall sub profiles with

variable sizes to detect the size of the queue. The size of the sub profiles varies according

to the distance from the camera. As the microcomputer systems operate sequentially, a

motion-detection operation is first applied, and then if the algorithm detects no motion,

the vehicle-detection operation is used to decide whether there is a queue or not. The

reason for applying motion-detection operation first is that the traffic scenes we analyzed

for queue detection are expected to contain vehicles, and in this case vehicle detection

mostly gives the positive result, while in reality there may not be any queue at all.

Therefore, by applying this scheme, the computation time is further reduced.

5.1. Motion detection operationThe method for motion detection used is based on differencing two consecutive

frames and applying noise removal operators. To reduce the amount of data and to

eliminate the effects of minor motions of the camera, the key regions of two frames, each

having at least three-pixel width, are selected along the image of the road. In this method,

a median filtering operation is first applied to the key region (profile) of each frame, and a

one-pixel wide profile is extracted. Then, the difference of two profiles is compared to a

threshold value to detect the motion. When there is a motion, the difference of the profiles

will be larger than the case when there is no motion. The process of motion detection is

shown in motion detection algorithm

The size of the profile for queue detection is an important factor as there might be

motion in some part of it, while there may not be motion in the other parts. Therefore, the

profile along the road is divided into a number of smaller profiles (sub profile) with

variable sizes. Our experiments show that the length of sub profile should be about the

length of the vehicle in order to assure that the operation of both vehicle and motion-

detection algorithms work accurately



5.2. Vehicle-detection operation

The vehicle-detection algorithm is only applied when the motion-detection

algorithm described earlier detects no motion. The vehicle-detection operation discussed

in Section 7 is applied on the profile of the unprocessed image. To implement the

algorithm in real time, it was decided that the vehicle-detection operation should only be

used in a sub profile, where we expect the queue will be extended (tail of the queue). This

procedure is shown in vehicle detection algorithm



6. MOTION DETECTION ALGORITHM

Block diagram

Fig(6.1). Motion detection algorithm


Median Filter

Dif = profile 1 – profile 0

Dif<T

h1

Frame 1 Frame 2

L L

Line profile 0 Line profile 1

Yes(No motion)No(motion)


Theory behind The profile along the road is divided into a number of smaller profiles (sub-

profiles)

The sizes of the sub-profiles are reduced by the distance from the front of the

camera.

Transformation causes the equal physical distances to unequal distances according

to the camera parameters.

Knowing coordinates of any 6 reference points of the real-world condition and t

coordinates of their corresponding images, the camera parameters (a11, a12…a33)

are calculated.

The operations are simplified for flat plane traffic scene - (Zo=0).

Above equation is used to reduce the sizes of the sub-profiles - each sub profile

represents an equal physical distance.

Number. of sub profiles depend on the resolution and accuracy required. The

length of sub profile should be about length of the vehicle - both the detection

algorithms work accurately then.



7. VEHICLE DETECTION ALGORITHM

As shown in Fig.7.1, to detect vehicles, windows with suitable sizes are placed

across each lane of the road. In this vehicle-detection approach, the SMED edge detector

is applied to each window, and the histogram of each window is processed by selecting

the dynamic left-limit value and a threshold value to detect vehicles.

Fig (7.1).The position and size of window



7.1 Left-limit selection programThe left-limit selection program selects a gray value from the histogram of the

window. When the window contains an object, the left limit of the histogram shifts

toward the maximum gray value, otherwise, it shifts toward the origin. For each 100-

frame sequence, the left limit for each window is the minimum of the left limits of the

previous sequence. In this approach, the minimum gray values of those windows, which

contain no object, is selected as a left limit. In case of queue detection, when all 100

frames detect a vehicle, the left limit is not changed.

Fig. (7.2). Histogram of the left limit of 200 frames after applying median

Filtering.



7.2Threshold selection programFor threshold selection, the number of edge points greater than the left-limit gray

value of each window is extracted for a large number of frames (200 frames). These

numbers are used to create the histogram. This histogram is smoothed by using a median

filter, and we expect to get two pecks in the resultant diagram one peck related to the

frames with vehicles and the other related to the frames without vehicles for that window.

However, as it is can be seen in Fig. (7.2), there are other numbers of edge points (30–40)

between the pecks 20 and 60, which are related to those vehicles which partially pass the

window. Where the sum of the entropy of the points above and below this point is

maximum. This point is selected as the threshold value for the next period.



8. TRAFFIC MOVEMENT AT JUNCTIONMeasuring traffic movements of vehicles at junctions such as number of vehicles

turning in a different direction (left, right, and straight) is very important for the analysis

of cross-section traffic conditions and adjusting traffic lights. Previous research work for

the traffic movement at the junction (TMJ) parameter is based on a full-frame approach,

which requires more computing power and, thus, is not suitable for real-time applications.

Use another method based on counting vehicles at the key regions of the junctions by

using the vehicle-detection method.

The first step to measure the TMJ parameters using the key region method is to

cover the boundary of the junction by a polygon in such a way that all the entry and exit

paths of the junction cross the polygon. However, the polygon should not cover the

pedestrian marked lines. This step is shown in Fig. (8.2).

The second step of the algorithm is to define a minimum numbers of key regions

inside the boundary of the polygon, covering the junction. These key regions are used for

detecting vehicles entering and exiting the junction, based on first vehicle- in first-

vehicle-out logic.

Following the application of the vehicle detection (described in Section III) on

each window, a status vector is created for each window in each frame. If a vehicle is

detected in a window, a “one” is inserted on its corresponding status vector, otherwise, a

“zero” is inserted. Now by analyzing the status vector of each window, the TMJ

parameters are calculated for each path of the junction. Since this process is also used for

counting number of vehicles, it does not increase the computation power requirement.

Fig (8.1).Result of TMJ Parameters.



Fig (8.2). The Polygon Covering the Junction.



9. RESULTS

The main queue parameters we were interested in identifying were the length of

the queue, the period of occurrence and the slope of the occurrence of the queue

behind the traffic lights.

To implement the algorithm in real-time, it was decided that the vehicle

detection operation should only be used in a sub-profile where we expect the queue

will be extended. The procedure

Fig (8.3).


N N N NY Y Y Y

Level>Th Level>Th Level>Th Level>Th

Motion detection Vehicle detection


QUEUE DETECTION

Fig (8.4).Queue Detection



Fig (8.5).Edge Detection



10. CONCLUSION

• Algorithm measuring basic queue parameters such as period of occurrence

between queues, the length and slope of occurrence has been discussed.

• The algorithm uses a recent technique by applying simple but effective

operations.

• In order to reduce computation time motion detection operation is applied on all

sub-profiles while the vehicle detection operation is only used when it is

necessary.

• The vehicle detection operation is a less sensitive edge-based technique. The

threshold selection is done dynamically to reduce the effects of variations of

lighting.

• The measurement algorithm has been applied to traffic scenes with different

lighting conditions.

• Queue length measurement showed 95% accuracy at maximum. Error is due to

objects located far from camera and can be reduced to some extent by reducing

the size of the sub-profiles.



11. REFERENCS

• M. Fathy and M. Y. Siyal, Senior Member, IEEE, “A Window-Based Image

Processing Technique for Quantitative and Qualitative Analysis of Road Traffic

Parameters”, IEEE TRANSACTIONS ON VEHICULARTECHNOLOGY, VOL. 47,

NO.4, NOVEMBER 1998, pp 1341 -1349.

• David Beymer, Philip McLauchlan, Benn Coifman and Jitendra Malik, “A Real-

time Computer Vision System for Measuring Traffic Parameters”, publication in

Transportation Research, December 1, 1998,pp 23-38.

• Rafael C.Gonzalez and Richard E.Woods, “Digital Image Processing”, Pearson

education, 3rd edition, pp 57 - 93.


real time image processing applied to traffic queue€¦ · web view2. signal & image...

Documents