Research On Human Abnormal Behavior Analysis Technology In Video Sequences

In recent years,with the rapid development of the Internet,big data and multimedia technology,video image resources are growing rapidly at an alarming rate.As an important information carrier,video images have entered all aspects of people's lives.Faced with such a huge amount of video image resources,how to efficiently and accurately identify human action and abnormal behaviors in video sequences has become a hot issue of concern.For the abnormal behavior recognition task in video,the main research is to analyze and process the human body motion in the video,extract the data features of the human motion and classify and identify according to the motion characteristics,so as to detect and alarm the abnormal behavior in the video.Traditional behavior recognition methods generally require manual extraction of features,and recognition accuracy is difficult to improve.In recent years,deep learning algorithms have been widely used in the field of computer vision,and computers can be used to automatically analyze data features in video.This topic focuses on the abnormal behavior of human body in video sequences,and conducts innovative research and exploration on abnormal actions such as robbery,punching car,falling,and fighting in complex backgrounds.Applying the deep learning method to the field of behavior recognition,an anomaly behavior recognition algorithm based on deep residual LSTM and dual-stream convolution fusion network model is constructed.In view of the abnormal behavior of human body in video,this paper mainly improves the original dual-stream convolution network and its derivative models from the following three aspects.Firstly,construct a dual-stream convolution structure based on C3 Dnet,and use the multiplicative cross-flow residual unidirectional connection method to integrate the network internally;Secondly,construct a deep residual LSTM module to fully learn the long-term motion information in the video;Finally,a scheme based on the dual-center loss function is proposed to optimize the network model.Finally,the subject is trained on the UCF101 and HMDB51 datasets,and then the model is migrated to the surveillance video dataset CASIA for abnormal behavior recognition,and the experimental results are analyzed to verify the effectiveness of the proposed algorithm.Figure 35;Table 9;Reference 49.