The Design And Implement Of Pedestrian Anomaly Detection System In Surveillance Video

Abnormal behavior detection is a valuable research topic in surveillance system. This thesis analyzed some challenges in surveillance system, researched its implement method, motion information subtract method and anomaly detection method of surveillance system, and designed a surveillance system for pedestrian anomaly detection. The main research works of this thesis can be described as follows:(1) Designed and implemented hardware architecture and software framework for anomaly detection in surveillance system. This architecture/framework, which can be divided into several levels, has the ability to support the expansion of hardware devices, such as cameras and CPUs, meanwhile it also support flexible switch of different algorithms, such as different pedestrian detect algorithms or anomaly detect algorithms. In order to make the system more easy to use and more intelligent, the system store video in the order of event which can boost the retrieval time, besides the system compressed surveillance video based on whether there is any motion in video, which saved storage spaces.(2)Researched the pedestrian detect and track method. This thesis on one hand realized a cascade pedestrian detection algorithm by combining the frame different method and pedestrian detect method using HOG feature, on the other hand this thesis researched and expanded the track method via spatial-temporal context learning, and realized multi-target tracking. In the end, this paper realized automatic tracking initialization and boosted the accuracy by combine detection information and tracking information.(3)For the indoor and outdoor scene of campus surveillance, this thesis realized rule-based anomaly detect method include cross line detection and border security. Besides, this thesis built pedestrian motion models based on long time analysis of surveillance video, presented an anomaly detection method via multi-motion model, and verified the effectiveness of the method.(4) The pedestrian density estimation algorithm was studied based on foreground pixels. When pedestrian density above a threshold, the abnormal behavior was detected using amplitude-based weight histogram of social force. Behavior detection in single people area and multi-people area were handled separately with different algorithms.