Research On Detection And Recognition Algorithm Of Human Abnormal Behavior In Public Area

The detection and recognition of abnormal human behavior under video surveillance has become a research hotspot in the field of public safety.However,video surveillance in real life has the shortcomings of mutual occlusion of human bodies and the difficulty of distinguishing similar abnormal behaviors,which leads to the problem of poor detection and recognition of abnormal human behaviors and low real-time performance.This paper researches on the detection and recognition algorithm of abnormal human behavior based on video surveillance,applies deep learning technology to the task of abnormal human behavior detection and recognition,and builds a two-level cascade network of abnormal behavior detection and recognition to realize the detection and recognition of human body in public areas.Effective detection and identification of abnormal behavior.The main research of this paper is as follows:1.Aiming at the problem of poor human abnormal behavior detection effect and low real-time performance caused by mutual occlusion of human bodies in complex scenes,a detection algorithm that uses generative confrontation network to detect abnormal human behaviors is proposed.The algorithm predicts the next frame of the video by generating a generative adversarial network,constructs a motion loss function to effectively extract the human motion characteristics in the video stream of complex scenes,and establishes an abnormal behavior decision function to achieve accurate positioning of abnormal behavior video frames,finally experiment on the self-built abnormal behavior data set.Experimental results show that the method can accurately detect the location of anomalous behavior video frames in complex scene situations with a detection accuracy of 98.0%,which is a5.7% improvement over the auto-encoder anomalous behavior detection algorithm;the detection speed of the video stream can reach 35 frames per second,which can realize real-time detection of abnormal human behavior.2.Aiming at the problem that it is difficult to distinguish similar abnormal behaviors in the abnormal behavior fragments located in the first-level network,which leads to the low accuracy of abnormal behavior recognition,a human abnormal behavior recognition algorithm based on the convolution of attention enhancement graph is proposed.By introducing bone length information flow and joint information flow data,constructing an adjacency matrix with global adaptability,assigning weights between unconnected nodes,fully learning the correlation between abnormal behaviors and joints,and adding time and channel attention mechanisms.Focus on information-rich joints,effectively extract distinguishing abnormal behavior features,use Softmax to identify the extracted abnormal behavior features,and finally conduct experiments on the abnormal behavior segments of the first-level network positioning.Experimental results show that this method can achieve91.2% recognition accuracy when similar abnormal behaviors are difficult to distinguish,which is 10.74% higher than that of spatio-temporal graph convolutional network behavior recognition algorithms.It is very similar to knife-holding and robbery behaviors.The recognition accuracy can be increased by 3% and 14% respectively;the recognition speed can achieve 25 frames per second,which can realize the accurate recognition of abnormal human behavior.