Abnormal Behavior Detection by Multi-SVM-Based Bayesian Network
Yufeng Chen, Guoyuan Liang, Ka Keung Lee, Yangsheng Xu · 2007
Automatic recognition of human actions is an important but difficult problem in the area of computer vision. In this paper, an novel approach is introduced to handle the problem. Firstly, human body is detected through the use of contour information and the body is tracked by the hybrid method. The most prominent features are searched by using the mean-shift method based on the body structure and the history motion image information. Finally, a learning method based on the multiple support vector machines is used to learn action types dynamically. We propose a method which integrate the Bayesian framework with the SVM method, which largely improves the recognition rate using historic information. Experiments show that our system can run in realtime for the detection of abnormal behaviors with limited information and produces robust result by making full use of historic motion information.