Study On The Vehicle Detection Based On Multiple Parts Model In 3D Space

Vehicle detection is an important part in intelligent transportation systems(ITSs). There are many vehicle detection algorithms at present, such as methods based on feature descriptors, classifiers, 3D modeling, et al. Though some advances have been made, many problems exist. For example, the performance is sensitive to the environment change and can be affected by object occlusion and merging. Considering the vehicle features in 3D space, the method based on multiple parts model in 3D space is used to fulfill vehicle detection.First of all, due to the projection between 2D image and 3D space, we propose to make a virtual inverse mapping plane in world space. Then the information on this plane is restored by using the inverse perspective mapping from space to image, which is called an inverse mapping image. The geometrical and size features of object are recovered in the inverse mapping image. To vehicles, according to the different light conditions between night and daytime, we choose pair-wise headlight as salient vehicle parts at night and license plate and rear-lamp at daytime. Different image process methods are used to detect pair-wise headlight at night and license plate and rear-lamp at daytime. The former is detected by geometrical feature, and the later are detected by color feature. After extracting spatial relationships between vehicle parts from plenty of vehicle samples, GMM of part-based spatial relationship is modeled by estimation Maximization(EM) method. Finally, if the probability like-hood value of detected parts in the inverse mapping image is bigger than the predetermined threshold value, those parts belong to one vehicle. So the detection result of multiple parts is vehicles’.The algorithm in this paper is tested in several traffic videos. The video frame frequency is 25 frames per second and the video size is 720*288. Experimental results show that the proposed approach is adaptable to weak light condition and vehicle occlusion. And it can avoid merging adjacent vehicles. Besides, it is applicable to vehicle detection with different type.