Study On Pattern Classification Of Human Behavior For Security Surveillance

With the increasingly serious security problem, the public places are generally installed video surveillance system; intelligent security surveillance has become a hot research topic in the field of computer vision. Intelligent video surveillance system can identify different objects and abnormal situation. At the same time, it can sound an alarm in the fastest and best way and provide useful information when danger is happen. In this way, it would effectively assistant people to deal with the emergencies in time. This paper focuses on moving target detection, feature extraction, behavior sequence segmentation and behavior pattern classification based on the "wireless mobile station security system" project.A thorough research on moving target detection algorithm is done from the perspective of theory and practical application. After comparing the advantages and disadvantages of the frame difference method, optical flow method and background subtraction method, extracting the moving target by using the method of Gaussians mixture model. For the disadvantages that the number of Gaussian distribution in Gaussians mixture model is fixed and existing waste of resources, this paper puts forward the adaptive method that choosing the number of the Gaussian distribution according to the changes of background, and optimizing the binary image, the result of detection is good.Most of the current behavior analysis algorithms are based on separate behavior pattern, according to this situation; the paper takes an in-depth study on the behavior sequence segmentation algorithm. This paper contrast and analyze the advantages and disadvantages of the human body outline features, regional features and space-time features, and on the basis of the method based on video sequence mutation point or discontinuity point, proposing a behavior sequence segmentation method which according to the changes of the intrinsic dimension of behavior space-time features. First of all, using the grid method to extract the space-time feature vector; then, using singular value decomposition(SVD) to estimate the intrinsic dimension of feature vector, to determine the corresponding low dimensional manifold of the data, and by detecting the projection error mutations on the manifold of the feature data, to realize behavior sequence segmentation. Finally, validated on the Weizmann database, better segmentation results have been achieved.Targeted to study basic definitions and basic algorithm of the Hidden Markov Model(HMM), effectively solve the three basic problems of Hidden Markov Model: assessment questions, decoding questions and learning problems. By HMM to verify the effectiveness of the behavior recognition based on KTH database, and then identifies the segmentation result of this paper, comparing with the recognition results of the Weizmann public database. The experiment shows the effectiveness of the segmentation and recognition algorithm.