Research On Abnormal Behavior Recognition Method Based On Multi-network Cascade Prediction

With the advent of the information age,the level of computer software and hardware has developed rapidly,and digital remote monitoring systems have become popular on a large scale,and the amount of collected video data has also increased sharply.A large amount of video data has brought a huge challenge to the monitoring management system with humans as the main body of observation.The intelligent behavior recognition system extracts the characteristic information in the image sequence through the computer,and constructs a behavior rule model to classify and recognize abnormal behaviors,which can improve the efficiency of security monitoring.However,the traditional end-to-end deep learning abnormal behavior recognition method is affected by the type and scale of the data set,the model has poor adaptive ability and limited ability to portray human behavior.Moreover,the definition of abnormal behavior generally depends on the scene,and accurate classification is difficult,resulting in poor realtime detection of abnormal behavior and low recognition rate of classification algorithms.This paper starts from how to extract the robustness of abnormal behavior recognition,in order to realize the rapid recognition of abnormal behavior of multiple people,combined with the regularity of pedestrian in normal movement,and the continuity of abnormal events and violations,proposes a method of abnormal behavior recognition based on multi network cascade prediction.The human body position is extracted by the instance segmentation model,and the skeleton tracking in video is completed by combining skeleton extraction with sparse optical flow.The dynamic skeleton information is predicted by bidirectional recursive codec network,and the abnormal skeleton score is compared with the threshold to judge the abnormal behavior.At the same time,in view of the problem that the participation of redundant skeleton information in the video acquisition affects the overall network computing efficiency,this paper proposes a joint feature method of skeleton joint information,which can more effectively represent human behavior and improve the accuracy of behavior recognition in complex background;In addition,in the abnormal behavior recognition module,by adding attention mechanism in the recurrent neural network,different weight values are matched for different parts of the human torso,which can make the neural network focus on the more important part of the abnormal behavior in the numerous input information.Finally,through the Shanghai Tech campus open data set and self-built data set,the experimental results show that this method has high detection accuracy in different scenarios and different types of abnormal behavior.