Research On Abnormal Behavior Real-time Detection Based On Video Feature

With the development of artificial intelligence technology,under the support of big data and cloud computing,intelligent video monitoring from computer vision technology is the forefront application development direction of networked video surveillance.The research of abnormal behavior detection under video monitoring is one of the key topics.Due to the growing demand for security in all fields of society,video monitoring equipment has been widely used in public places such as junctions,shops,banks,metro stations and schools to ensure social security.The traditional video analysis mainly depends on human and takes a lot of manpower and resources,but the advantages of digital and networked video monitoring system is more and more obvious compared with the traditional system.Therefore,based on the theory and practice,this article makes a new exploration of the abnormal behavior detection in video monitoring.In this paper,the method of global abnormal behavior detection based on double sparse representation is introduced firstly.In this method,the video frame is first divided into several regions of the same size,and a container is formed along the time axis.Then,for each sub region in the spatio-temporal container,the hybrid optical flow histogram is calculated.Finally,the hybrid optical flow histogram is used into sparse reconstruction,and enough training samples are selected to establish the initial normal behavior dictionary and abnormal behavior dictionary,and constantly updated in the training process,after a complete abnormal dictionary is established,the coming sample will have two sparse processes,and then the global abnormal behavior is detected by fuzzy integral.In this paper,the method of local abnormal behavior detection based on online weighted clustering is introduced.Firstly,the salient object detection based on Harris corner is applied to the input video stream,that is,detecting the active area of the crowd based on the motion information of the object and the extraction of corner feature,and Bayesian method and saliency diffusion are used to optimize active region.Then,the features of the hybrid optical flow histogram are extracted in the region of interest.Finally,the classification of test video is realized by using online weighted clustering and multi target tracking is used into optimization.Experiments on public UMN dataset and UCSD dataset show that the global and local abnormal behavior detection methods are proposed in this paper can not only avoid the influence of light and human occlusion,but also have higher performance and precision than the previous methods indetection performance and evaluation criteria.