Research On Crowd Feature Perception And Abnormal Detection In The Video

Accompanied by the progress of urbanization and the increase of population mobility, more and more attention has been paid to the social public security issues. Video abnormal event detection which is one of key research of intelligent video surveillance application can contribute to probing the crimes, accidents and terrorist attacks. It is important for the the maintenance of social public security. This paper does the deep research in two aspects as feature perception and model abnormal detection. According to the problems existed, the corresponding optimization algorithm has been proposed. The contents summarized as following:In the crowd abnormal event detection, the crowd feature is the sociality performance of all the individuals in the crowd. As the individuals not only present the sociality but also the randomness, it needs to coordinate the opposition of the sociality and randomness in the process of crowd feature perception. So this paper proposes a crowd feature perception algorithm based on the compound prior sparse coding. It builds the statistical characterization of the sociality and randomness, proposes the compound sparse prior description, and considers the individual behavior consistent with the compound prior of sociality and randomness. So it can enhance the adaptability of crowd feature perception and make the feature better suit for the complicated crowd scene. By optical feature extraction and compound prior sparse coding, this method has achieved the effectively and steadily feature description. Moreover, it combines the iHMM statistic model to detect the crowd abnormal event. The experiment on the UMN datasets and simulated datasets shows this algorithm has a high level of accuracy and a wide applicability in crowd abnormal event detection.In the video abnormal detection based on the statistical models, the ability to fit the observation data depends on the statistical model. So it is important to optimize the statistical models to avoid the under-fitting or over-fitting problem. The complexity of spatial-temporal coupling in the video abnormal event determines that the optimization of statistical models doesn't only depend on the states dimension of model, but also depend on the structure of the model. So this paper proposes a video abnormal detection algorithm based on joint spatial-temporal model. This model is based on DDP-HMM and builds undirected connections between LDA topic features and HMM states. So it can optimize both of the states number and the model structure to increase the ability of model description. The algorithm is composed of two stages as mode training and abnormal detection. As the joint spatial-temporal model is a directed cyclic graph model, it uses the method of weighted tree to learn model parameters. First converts the model to two spanning trees as iHMM-LDA and LDA-iHMM structure and learn each parameter of the structure, then weight the parameter to optimize the joint spatial-temporal model. When in the stage of abnormal detection, it calculates and compares the log likelihood function of the train videos and test videos by forward algorithm and loop belief propagation algorithm. The experiment on the public data sets UCSD and UMN verifies the excellent detection performance of the joint spatial-temporal in the abnormal event detection.