Research On Action Recognition Based On Feature And Deep Learning

Human action recognition is a hot research direction in machine vision,video-based human action recognition is becoming more and more popular.It has broad application prospects in such areas as video surveillance,human-computer interaction and behavior analysis.In this paper,the human action recognition and dynamic video is deeply studied.First,aiming at the video instability,the improved Steady Flow algorithm and the adaptive multipath optimization algorithm is proposed,three human body detection network are trained to reduce the background interference to action recognition.Second,the optimization of features is studied,the feature positive skew distribution optimization algorithm and action recognition based on feature positive skew distribution are proposed.Finally,a new action recognition deep learning network model is studied and used to recognize human action.The main research of the thesis is as follows:1.Because the instability of video can interfere the accuracy in action recognition,the Steady Flow algorithm and the Path Optimization algorithm in video stabilization is studied.The improved FAST feature points is used in Steady Flow and the adaptive multipath optimization algorithm is proposed.Aiming at the problem of scene motion and background interfere to action recognition,the background is fixed and the human body recognition network is used to identify the human body.2.Aiming at the redundancy of action features,the feature positive skew distribution optimization algorithm is proposed by studying the distribution of action features.Based on the optimization algorithm,an action recognition method based on feature positive skew distribution is also proposed.The method use Dense Trajectories feature and Time-space Interest Points feature,and combined with the feature positive skew distribution optimization,Fisher Vector,SVM classifier,feature fusion to realize action recognition.Using the stable and human-detected data set to experiment,the recognition accuracy is improved.3.By the study of Convolutional 3D,Residual Network,Dense Network and Residual Dense Block,Convolutional 3D Residual Dense Block(C3D-RDB)is proposed,which combined with the residual learning,dense connect structure,and it can fulfill Convolutional 3D.Furthermore,the 3D Convolutional Residual Dense Net(C3D-RDN)is constructed based on C3D-RDB,Using the C3D-RDN to recognize action,which achieved a good results.