Study On Detection Approach Of Abnormal Events In Surveillance Video

At present, numerous digital video surveillance systems fail to help security people to detect abnormal events timely and automatically in case of no guard, as a result, the recorded video data can be used as evidence just after events happened. Introduced to deal with this problem, intelligent surveillance system can detect the abnormal events and report to security human in time from redundant video data. Just based on this purpose, a testing program was designed to detect abnormal events.The detection of two kinds of abnormal events named wandering and battering were realized in the testing program. In special scene such as bank or parking, wandering could be considered to some extent as abnormal event, showing by multiple changed times of motion direction of targets in certain time. Thus, based on the motion trajectories of targets obtained from the scene, it is capable of finding out whether the wandering event is abnormal or not by analyzing the directional features.According to the divisibility between wandering events and others in vision, it is possible to dig out the motion features which could differentiate the abnormal events from the others. In the testing program, it extracts four types of motion activity features, including motion intensity, motion angle histogram, run-length and intensity ratio from fragments which defined as one second long video. By finding out those thresholds from the video data, the program could classify the video into two kinds which named battering and normal event to detect the abnormal events from real-time video data automatically.As the results show, it is practically to detect the wandering events according to the motion trajectory analysis approach, in which the foreground targets could be segmented with good effect from video sequence using local processing method in the interesting regions. While in case of complex scenes, the proposed method using motion activity features is a good choice to detect the battering events and it is proved to have high recall and precision.