The Crowded Abnormal Event Detection In Video Surveillance

With the development of modern social economy, the speeding up of urbanization, large-scale activities have become increasingly frequent, and the related accidents are common. Therefore, how to monitor the activities management is one of the hot issues to be solved in the process of urbanization.Intelligent video surveillance network is the technology to use computer technology to analyze the videoautomatically. As an effective way of public security, intelligent video surveillance system is more and more get the favour of people. Andabnormal behavior detection technology in intelligent video surveillance is becoming more and more attention in the field of intelligent video surveillance. With the computer technology, the analysis of the mass video data in the video surveillance network, compared to the traditional manual analysis, can greatly reduce their worktime and due to great economic benefits for the society. Also, with the compute power of our devices, our surveillance network can solve the problem that the manual analysis can’t be real-time. In this paper, we have some new ideas about the abnormal behavior detection. The main contributions of this thesis are summarized as follows:(1)A new online dictionary learning method is presented to replace the state of art methods like MOD and KSVD. We cast the dictionary learning problem as the optimization of a smooth nonconvet objective function over a convex set, minimizing the expected cost when the training set size goes to infinity. In each step, the quadratic surrogate function of the empirical cost over the set of constraints can be solved with LARS. Compared with the traditional methods, our method can more quickly reach the convergence and get a low storage cost.(2)A new model name sparse combinations is used for the detection of abnormal events. In the training stage, this model updates the combinations to make each combination can represent the video data with a low reconstruction cost as much as possible. The optimization problems can be casted to be a simple quadratic optimization problem, to reduce the traning time. Also in the detection stage, our method only needs to loop the sparse combinations to see that whether any reconstruction cost is under the threshold. If yes, then we judge that it is a normal event,Otherwise, it is an abnormal event. But with the traditional method, we need to search the coefficient space to get the minimum reconstruction error to judge whether it is beyond the threshold. It cost too much compute power and due the low fps of the system.(3)A Mixture of Dynamic Textures is presented to model the video data. Compared to the only one Dynamice Texture, the MDTs use the linear weighting of each DT can model the video data more precisely, especially for the complex vaiances video data.(4) A new abnormal event detection method based on the fuse of the temporal detection and spatial detection is used in our paper. However, the traditional methods only take account of only one of the temporal abnormal event and spatial abnormal event. In this paper, we use the temporal models to compute the temporal anomaly maps. Meanwhile, use the saliency algorithms to compute the spatial anomaly maps. Then, we fuse each map to get the globally consistent anomaly maps and use this map to judge whether the abnormal events happen.