Stereo Vision based On-road Vehicle Detection under Illumination

Changing Conditions using Self Quotient Image

Jonghwan Kim, Chung-Hee Lee and Young-Chul Lim

Daegu Gyeongbuk Institute of Science & Technology, 50-1 Sang0Ri, Hyeonpung-Myeon, Dalseong-Gun, Daegu, Korea

Keywords: Stereo Vision, Vehicle Detection, Vehicle Recognition, Self Quotient Image, Illumination Equalization.

Abstract: Today the many of automotive research groups study how to reduce vehicle accidents. For this reason, they

have been developing the advanced driver assistance system (ADAS). In ADAS, the various sensors are

used for recognizing the driving situations. For example, there are supersonic wave sensors and radar

sensors and so on. In particular, in computer vision research groups, the vision sensors (ex. CCD, IR) are

used for this. But it has some difficult problems because the vehicles are mainly driven in outdoors. The

images captured by outdoors have various illumination conditions due to weather. It makes difficulty to

detecting vehicles in images. In this paper, we introduce the vehicle detection method when the input

images of system have illumination changes. We use the self quotient image (SQI) algorithm for

illumination equalization. But SQI algorithm produces many false positive results. So we eliminate the

false-positive results using stereo vision technique. In main section, we explain this method in detail. And

we prove the proposed method has superior performance than existing systems using experiments.

1 INTRODUCTION

Every minute on average, at least one person dies in

vehicle accidents. In addition, specific information

hasn’t been known, vehicle accidents make many

casualties or property damages. According to this,

the research groups associated with automotive

industries are researching and developing various

methods for reducing vehicle accidents. One of these

is the advanced driver assistance system (ADAS).

The ADAS notifies on-road situations to the drivers

and helps their safety driving. For this, ADAS

should be able to detect the obstacles and

information of road. So, in various fields, the

researchers are developing the objects detection

system for ADAS. For example, there are the smart

parking assist system (SPAS) using ultrasonic wave

sensor, advanced smart cruise control(ASCC) using

radar sensor, night vision using infra-red(IR) sensor

and so on. In recent years, especially, the object

detecting methods using vision sensors (e.g. CCD,

CMOS) are being studied by the computer vision

researchers. But the object detecting technique in

image has some difficulties. The images from vision

sensor have many of the environmental variables

according to locations where images are taken.

Especially, vehicles are mainly driven on outdoors

and the environment of outdoor has various

illumination changes. Figure 1 shows the

illumination changing effects on vehicle’s

appearance. The images are taken continuously

when the vehicle passes under the bridge. The

vehicle appears differently by shadows and

reflections.

Figure 1: Illumination change effect.

In this paper, we introduce a vehicle detection

method when there are any changes of illumination

in image. We use captured images from driving

vehicle and the images hold the forward looking of

vehicle. Stereo camera equipped with CCD sensors

581

Kim J., Lee C. and Lim Y..

Stereo Vision based On-road Vehicle Detection under Illumination Changing Conditions using Self Quotient Image.

DOI: 10.5220/0004164605810584

In Proceedings of the 9th International Conference on Informatics in Control, Automation and Robotics (IVC&ITS-2012), pages 581-584

ISBN: 978-989-8565-22-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

was used in the experiment and the output image

size of camera is VGA (640480). First of all, the

self quotient image (SQI) algorithm is applied to the

output image of camera. SQI algorithm has been

mainly used in the field of face recognition and

effects on illumination normalization. But SQI

decreases the contrast sensitivity of images and

decreases classification ability of the classifier also.

It makes that the classifier outputs many false-

positive results. To solve this drawback, we propose

the stereo vision technique. We set the region of

interests (ROI) using depth map from stereo camera

output. The detection window of classifier is search

on the ROI only and the false-positive results of

other region (out of ROI) are limited.

The paper is organized as follows: Section 2

analyzes the related approaches, we will explain the

SQI algorithm and stereo vision based vehicle

detection method in this section. Section 3 shows

our proposed method. The experiment results and

performance comparison are presented in Section 4.

Finally, the conclusion is described in Section 5.

2 RELATED APPROACHES

2.1 Self Quotient Image

In this subsection, we explain the Self Quotient

Image (SQI) briefly. As mentioned in the previous

section, SQI has been used in face recognition field

mainly and effects on illumination normalization.

For illumination normalization, above all, the light

analysis is needed. The light analysis is widely used

in the Retinex algorithm. The Retinex algorithm is

compensation method for images degraded by light

changing. It is based on the reflectance illumination

model like equation (1).

  

(1)

Where  is the image,  is the reflectance of the

scene and  is lighting. The light is considered as the

low frequency component of image because it

mainly affects plane region of objects. So  is

represented as the following equation.

    

(2)

We consider the F is a Gaussian filter and * is

convolution operation. According to the equation

above,  of (1) can be expressed again like equation

(3).

 









  

(3)

Using the reflectance component  , the SQI is

defined as follows:

 









  

(4)

Where Q is self quotient image (SQI), 



is smoothing

result of input image , F is smoothing filter. Figure

2 is SQI algorithm results from Figure 1.

Figure 2: SQI results.

Like Figure 2, the results of SQI algorithm

similar to high-pass filter results. Those are same as

if they were edge-highlighted images. And we can

see the illumination changes of plane region are

removed. Especially, SQI is very effective to remove

the shadow region.

2.2 Vehicle Detection in Image

Figure 3: Stereo vision based vehicle detection.

Figure 3 shows the system flow chart that explains

the stereo vision based vehicle detection method.

This system is divided into two main parts. First part

is ROI setting. Second part is classifier learning. In

this paper, we introduce this method briefly because

this method was fully introduced in our previous

papers.

ICINCO 2012 - 9th International Conference on Informatics in Control, Automation and Robotics

582

2.2.1 ROI Setting using Stereo Vision

Like Figure 3, the stereo matching part gets two

images from stereo camera and computes the depth

information. We calculate the matching cost using

normalized cross correlation (NCC) and optimize

the matched costs using hierarchical belief

propagation (HBP). From this depth map, we

estimate the road region using U-V disparity

calculating method. From this information, we can

eliminate the road region and set the remaining blob

to ROI. Needless to say, because of depth estimation,

we can know the distance from blobs.

2.2.2 Classifier Learning

We use the AdaBoost classifier and Haar-like

feature for vehicle detection. The AdaBoost

classifier is learned by our vehicle image database.

This database is made-up by positive images and

negative images. The positive images have vehicle’s

appearances and the negative images have

backgrounds of on-road. To get better performance,

we made the image database with our hand but do

not use opened one. We set the cascade of AdaBoost

classifier to 14 levels and use the Haar-like features

that specialized in vehicles.

3 PROPOSED METHOD

Figure 4: Proposed method.

Figure 4 shows our proposed method. In this method,

the AdaBoost classifier is learned by image database

that processed SQI algorithm. And the detection

window of classifier searches on SQI applied input

images. As we explained, the SQI has drawbacks of

making many false-positives. To overcome this, we

set the ROI using stereo vision technique like

subsection 2.2.1. In next section, experiments, we

evaluate our proposed method.

4 EXPERIMENTS

Figure 5: Our experimental vehicle and system.

Figure 5 shows our experimental environments. We

got the database and test images from stereo camera

mounted on experimental vehicle. For real-time

processing, the stereo matching part is implemented

in hardware (FPGA) and other part is implemented

in PC. The used PC has Intel® core™ i7 CPU 2.67

GHz, RAM 4.0GB and Microsoft Visual Studio

2010. For evaluation tests, we use the images below.

Figure 6: Test scenes.

Like Figure 6, the test image sets have

illumination changing condition. The scene #1 has

40 frames, the scene #2 has 150 frames, the scene #3

has 100 frames and the scene #4 has 100 frames.

Total 390 frames are used for our test.

Figure 7: Example of ROI effects.

In Figure 7, the left result have many false-

positives. Like the right result, however, our method

removes the false-positives using ROI.

Stereo Vision based On-road Vehicle Detection under Illumination Changing Conditions using Self QQuotient Image

583

Figure 8: Example of SQI effects.

In Figure 8, the performance of proposed method

is better than the conventional one. This

experimental result can be seen in YouTube web site

(http://www.youtube.com/watch?v=sUwWvBMLPh

s).

Table 1: Evaluation results of proposed method.

Scene

F-measure

Stereo

0.786

0.712

0.669

0.754

Stereo

+SQI

0.893

0.801

0.692

0.796

   

    

  

(5)

Table 1 represents comparison of performance. In all

scenes, our method has better performance than the

conventional method. In general shadow region, the

performance of our method is superior. In

completely dark place like scene #3, however, the

both classifiers cannot detect the vehicles well.

5 CONCLUSIONS

We have proposed a vehicle detection method under

illumination changing conditions. In various

illumination conditions, we could see our method

has better performance than existing one. In outdoor

image, the vehicle detection must consider the

illumination effects. Our proposed method

contributes to improving performance of vehicle

detection. If this method has further improving, the

intelligent vehicle technology will be more perfect.

In this paper, we could not make accurate

performance evaluation because of lack of test

scenes. Though more tests, we will get the more

accurate results and continue to improve this method.

ACKNOWLEDGEMENTS

This work was supported by the Daegu Gyeongbuk

Institute of Science and Technology R&D Program

of the Ministry of Education, Science and

Technology Korea (11-IT-02).

REFERENCES

Sun Z., Gebis G., Miller R., 2006. On-Road Vehicle

Detection : A Review. IEEE Transactions on Pattern

Analysis and Machine Intelligence, Vol. 28, No. 5.

Wang H., Li S. Z., 2004. Face recognition under varying

lighting conditions using self quotient image. Proc. 6

IEEE Int. Conf. On Automatic Face and Gesture

Recognition (FGR’04).

Kim J., Lee C., Lim Y., Kwon S., Park C., 2011. Stereo

Vision-Based Improving Cascade Classifier Learning

for Vehicle Detection. Lecture Notes in Computer

Science, Advances in Visual Computing, Vol. 6939, pp.

387-397.

Kim J., Lee C., Lim Y., Kwon S., 2011. Optical Sensor-

based Object Detection for Autonomous Robots. Proc.

IEEE Int. Conf. On Ubiquitous Robots and Ambient

Intelligence (URAI 2011). ,

Kim J., Lee C., Lim Y., Kwon S., Park C., 2012.

Multithreaded Programming for Real-time Vehicle

Detection. Proc. Int. Conf. On Electronics,

Information and Communication (ICEIC 2012), pp.

194-195.

Kwon S., Lee C., Lim Y., Lee J., 2010. A Sliced

synchronous iteration architecture for real-time global

stereo matching. Proc. of SPIE-IS&T Electronic

Imaging, Vol. 7543, No. 754312-1.

Lee C., Lim Y., Kwon S., Lee J., 2011. Stereo Vision-

based Vehicle detection using a road feature and

disparity histogram. Optical Engineering Vol. 50, No.

Viola P., Jones M., 2011. Rapid object detection using a

boosted cascade of simple features. Computer Vision

and Pattern Recognition, pp. 511-518 (CVPR ’01).

Van Rijsbergen C. J., 1979. Information Retrieval.

Butterworth-Heinemann, London.

ICINCO 2012 - 9th International Conference on Informatics in Control, Automation and Robotics

584