Human Abnormal Behavior Recognition In Video Surveillance

With the rapid development of information technology gradually to the intelligent and digital as well as the improvement of people’s security consciousness, intelligent video surveillance is playing an increasingly important role. Among them, the abnormal behavior understanding and recognition of Human body is the advanced stage of the intelligent video surveillance, it realizes the understanding and recognition of individual action, between people, between people and environment by feature extraction,description and analysis of the foreground target to achieve the purpose of monitoring and early warning of emergencies.Paper studies and designs the algorithm process of Human abnormal behavior recognition in video surveillance by analyzing the video images. The human abnormal behavior recognition algorithms will be divided into human moving target segmentation,human behavior recognition based on template matching, human target tracking under complex background, human abnormal behavior recognition based on the trajectory.Human moving target segmentation is the foundation of abnormal behavior recognition. According to different scenarios, this paper uses three frame differential method, background subtraction method and mixture gaussian method for simulation,analyzing the performance of each algorithm from the perspective of robustness and real-time and determining the appropriate method of moving target detection of the experimental scenarios.Human behavior recognition based on template matching is founded upon accurate moving target detection. Firstly, with a series steps of extension, modification and normalization on the basis of traditional Hu invariant moments, this paper constructs weighted Hu moment using the minimum variance as the criterion. Compared with the traditional Hu moment, the weighted Hu moment gets more image information, and with scaling invariance in the discrete case. Besides, the weighted Hu moment fully considers the contribution values of each Hu moment invariants, making a certain improvement of the recognition performance. In addition, this paper puts forward a method of self-adaption features fusion of human behavior recognition based on weighted Hu moment and HOG, which can adjust the parameters of feature fusion according to the noise of the scene. Comparing with the sigle HOG or weighted Hu moment, this new method has wide scope, high noise tolerance and stable recognition rate.In human target tracking under complex background,, this paper introduces multi-feature fusion MeanShift tracking algorithm based on the traditional MeanShift tracking algorithm. Aimed at the poor tracking performance of the traditional MeanShift tracking algorithm when the gray scale of target and background is adjacent, in order toachieve accurate position and stable tracking, this paper introduces a improved MeanShift tracking algorithm method based on feature fusion of gray scale and HOG which combines the gray scale feature with the regional feature.Human abnormal behavior recognition based on the trajectory is affected by stable tracking. It realizes cross-line,wandering, clustering and spread of typical human abnormal behavior detection through the analysis of the trajectory,with simple algorithm,good real-time and high recognition rate.In order to improve the robustness and accuracy of the algorithm, this paper studies multiple key technology of human abnormal behavior recognition and obtains good experiment result.