Research Ofvehicle Detection For On-board Vision Systems

Vehicle detection is an important work in driver assistance system. Moving vehicles threat drivers’ safety most in real traffic scene. If the system can detect vehicles around the driver accurately and then warn the driver, traffic accidents will decrease largely. Cars change a lot in color, shape, perspective and light intensity when moving and thus make vehicle detection a challenging work.This paper researches on vehicle detection system based on monocular vision and presents a vehicle detection method including vehicle hypothesis generation and vehicle hypothesis verification. In the vehicle hypothesis generation step, shadows under vehicles are used to generate vehicle hypotheses. Shadow under vehicles is an important character of moving vehicles on the road. Gray value of this area is greater than another area on the road, because light cannot go through vehicles. This paper proposes using the vanishing point location to locate the region of interest(ROI). Firstly, the vanishing point is calculated and used to locate ROI. Then edge information is used to extract the free driving space. By counting histogram of the free driving space, the threshold for the shadow extraction can be obtained automatically.In this way, shadow can be extracted in different light conditions.The geometric models of vehicle images are analyzed and then used to filter shadows according to the width and position of the vehicle in the image. Finally, the hypothesis is generated based on the shadows. In the vehicle hypothesis verification step, this paper present two kind of histogram of oriented gradients(HOG) descriptors to extract vehicle features. There exist a large number of horizontal and vertical gradientsin thevehicle imagebecause of the structure of vehicles. This paper combines the two HOG feature vectors as the final feature vector. AdaBoost classifier is trained here. Finally, the hypothesis verification is generated by using the trained AdaBoost classifier. The extraction procedure of the combined HOG features costs less time while ensuring a high recognition rate.Experimentsarecarried out in MATLAB platform. The GTI vehicle database is used in the experimentsand the recognition rate can reach up to97.82%. Experiment results show that the proposed method has high robustness, accuracy and can meet the real-time requirement.