Improved PAM-based traffic behavior recognition using trajectory-wise features

Recently CCTV-based behavior recognition have gained considerable attention in the transportation surveillance systems to identify normalities, such as traffic jams, accidents, and dangerous driving. An improved method is presented in this paper for the traffic behavior surveillance system by discovering more highly specific features based on the trajectory information. The multiple sparse feature comprising the object location, moving direction, speed, and appearance time length obtained from the moving object detection and tracking stage is modeled by the Pachinko Allocation Model. This hierarchical probabilistic model captures the correlation among the traffic activities and behaviors through the sparse features as the visual words. In the classification phase, the Support Vector Machine constructed from Decision Tree Architecture is utilized. Compared with existing methods, the proposed method outperforms 3-8% approximately in overall classification accuracy.