Research On Multi-view Human Action Recognition Method Based On Deep Learning

With the continuous development of intelligent video surveillance technology,there are more and more application requirements for analyzing video images of people appearing to obtain characteristic information such as the action categories of people in the video,such as anti-terrorism,epidemic prevention and motion analysis.These applications require the system to automatically identify and analyze human motion videos to determine the action categories,identities and other information of people in the videos.At present,the commonly used algorithms for action recognition on two-dimensional video images of human motion collected by a single camera are often affected by environmental occlusion,human self-occlusion,and viewing angle deviation.One of the methods to improve the recognition accuracy is to use the video images of multiple cameras from different viewing angles in the same area for fusion analysis.This paper improves a multi-view human action recognition method based on 3D residual network 3D Res Net and long short-term memory network LSTM for human action recognition of multi-view camera video images.On this basis,a multi-view human action recognition and comprehensive analysis system is constructed to realize comprehensive analysis of various types of human body information in the video.The full text works as follows:First,in view of the problems of poor multi-view feature fusion and low recognition accuracy in the existing multi-view action recognition algorithms,a 3D Res Net-LSTM algorithm is improved,which automatically learns multi-view action sequences through a 3D residual neural network.Then,in order to reduce the influence of redundant background in the video on the accuracy of action recognition,the video frame is preprocessed,and the edge detection algorithm is used to reduce the influence of the redundant background in the video.The human body contour is extracted and cropped.The results indicate that this approach can decrease the computation cost and enhance the precision of the model.Thirdly,a multi-view comprehensive monitoring and analysis system is constructed,which is based on the abovementioned multi-view human actions.The identification method further improves the nonfrontal human identification method and combines the multi-camera human positioning method based on Open Pose.Through the experimental verification on the NTU-RGB+D 120 dataset,the results show that the multi-view action recognition rate of the algorithm reaches 83.2%.The constructed multi-view comprehensive monitoring and analysis system can realize the effective analysis and display of human body movement category and human body identity.