With the widely application of high-definition cameras in intelligent transportation system and the rapid development of image processing technology, image-based vehicle detection technology has become one of the most important vehicle detection techniques. Currently the image-based vehicle detection methods are extensively studied and their detection accuracies continue to increase. However, vehicle detection methods in urban traffic scene face some challenges: 1) vehicle occlusion. The traffic congestion causes vehicle occlusion which leads to inaccurate vehicle positioning. 2) non-vehicle object interference. In urban traffic scene, especially at the intersection, there are lots of non-vehicle objects, such as pedestrians, non-motor vehicle, etc. These objects degrade the performance of vehicle detection methods. 3) multi-scale vehicles. The image size, clarity and viewing angle of a vehicle are different when the vehicle is at different distances from the camera. We call this phenomenon a multi-scale problem, which has posed a challenge for vehicle detection due to the time varying nature of vehicle features. To circumvent above mentioned problems, in this thesis we propose several vehicle detection methods based on And-Or Graph (AOG). These methods can deal with the problems of vehicle occlusion, non-vehicle object interference and multi-scale vehicles,because the AOG utilizes compositional structure for vehicle representation and integrates structural, geometric, appearance, and probabilistic information for vehicle detection. The main contributions of this thesis are listed as follows: 1. To improve the performance of vehicle detection methods in processing vehicle occlusion, we develop a novel vehicle detection method called vAOG. In our method we construct a vehicle AOG to represent vehicle objects in urban traffic condition. Then we apply bottom-up inference to pursue vehicles in a traffic image. During the construction of vehicle AOG, we select the window and taillight as the vehicle features due to that they are often visible when vehicle occlusion occurs. Furthermore, we hierarchically decompose these features into smaller part-features. If a vehicle is occluded, there are still several part-features visible for detection. The constructed vehicle AOG can well describe the vehicles in occlusion condition. Finally, based on the vehicle AOG model we use bottom-up inference and combine detected part-features to detect the vehicles. The qua...
修改评论