Research On Feature Extraction And Recognition Technology Of Crowd Abnormal Behavior Based On Security Video

With the rapidly increasing demands from intelligent applications of Closed-Circuit Television(CCTV)in security industries,the technologies of crowd abnormal behavior feature extraction and recognition have become a hot research topic in the aspect of crowd monitoring and management.The interaction information between pedestrians cannot be fully extracted due to occlusions and shadowing among pedestrians,and also there is no mature an algorithm that promotes the low-level features of images to the high-level semantic features with human prior knowledge.Therefore,the recognition rate of crowd behaviors are still very low.In order to solve these problems,this research has investigated an automatic online crowd anomaly detection model by exploring a novel compound image descriptor from video streams and the training of a deep learning network based on these descriptor instances.In this paper,the work reported has focused on:1.analyzing the cause of the Gaussian Mixture Model(GMM)update rate slower,and proposes to the adaptive frame difference(TD)into GMM algorithm,and it has two improvements: 1)Adaptive threshold was added to improve the temporal difference to eliminate redundant noise caused by illumination change;2)The improved temporal difference method(TD)is incorporated into the GMM model,such that the extracted foreground areas can effectively remove the irrelevant backgrounds for reducing the computational costs in the subsequent processing of crowd Tracking.2.integrating a rich image descriptor(named spatio-temporal feature descriptor – STFD)of the scene independent,and it has three improvements: 1)The problem of information loss among subgroups caused by the CF(Coherent filtering)algorithm,and the segmentation algorithm(ST)of spatio-temporal cuboids is proposed,and The ST algorithm solves the problem of over-segmentation of the group and makes up for the interaction information among groups;2)adding a new local density attribute to calculate the compact degree of crowd in spatial distribution;3)improved the four attributes(namely,collectiveness,stability,conflict and density)of groups,and these interactive information of the crowd was integrated into STFD instances.The STFD not only improves the performance of feature extraction,but also improves the recognition rate of 4.75%.3.proposing an abnormal behavior recognition model based on convolution neural network model(CNN),and inputting generated STFD descriptor instances and original image into CNN to obtain the CNN model with high-level semantic features.The CNN model not only combines the static scene features and dynamic movement features of crowd,which improves the recognition rate of crowd behaviors.Lastly,based on the previous investigations,the thesis devised an automatic online crowd anomaly detection prototype system,evaluated the various functional modules of the system.This thesis integrated the CNN anomaly detection system with traditional anomaly detection system in accuracy and performance is compared and verified.The experiment results show that the CNN recognition model achieves high recognition rate and can be used in complex scenes.