Research On Human Skeleton Action Recognition Algorithm Based On Deep Reinforcement Learning

At present,human skeleton action recognition has far-reaching research value and significance in video retrieval,video abnormal behavior monitoring,human assisted training and correction.The human action recognition is essentially a video recognition task,but the redundancy of human action video sequence is inevitable.To solve the problem of the human skeleton action keyframe selection in a long time range,this paper proposes a reinforcement learning framework based on diversity reward and representativeness reward mechanism,which is a convolutional neural network for skeleton keyframe extraction in time dimension.The overall network framework used in this paper firstly selects the most diverse and representative skeleton frame data in time dimension by the keyframe selection network,and then inputs them into the powerful spatial-temporal feature extractor for human skeleton action label recognition.The diversity reward and representativeness reward mechanism designed by this paper has good compatibility with powerful feature extractors: two-stream adaptive graph convolutional networks(2s-AGCN)and multiscale graph convolution and unified spatial-temporal networks(MS-G3D).Good experimental results have been obtained on three ske leton data sets.The main work is as follows:(1)The human skeleton action recognition method based on diversity reward(DDRL-GCN)first models the video diversity frame selection process of video as a discrete Markov decision process.Simultaneously,this method abandons the previous practice of relying on the total category number of correctly identified actions of the current selected frame obtained by interaction with GCN as a reward.In video frames,key frames often occur in a few scattered frames in the fast-changing part of the action,and these frames are crucial to the integrity of the action construction.Therefore,in this paper,the variant of cosine distance is used as the similarity measurement in dex between frames in video,which is used to capture those frames with large difference in direction from adjacent frames to supplement the diversity of selected frame sequence..Finally,this paper input the selected diversity frame into the GCN-based spatio-temporal feature model selector containing MS-G3 D and 2s-AGCN models for accuracy recognition.To test the comprehensive performance of the algorithm,this paper compares DDRL-GCN with other cutting-edge algorithms in model parameters,GFLOPs and algorithm accuracy,and verifies the feasibility and effectiveness of DDRL-GCN.(2)The human skeleton action recognition method based on representativeness reward mechanism(RDRL-GCN)introduces Euclidean distance as the similarity measurement index between frames in video.The index continuously encourages the agent to select those representative frames whose distance difference from adjacent selected frames in the feature space is as large as possible to construct more continuous actions.Simultaneously,the following problems still exist in the selection process of DDRL-GCN frames: The chronology of selected frames is difficult to ensure.In the process of key frame selection,the action initial frame which is crucial to action continuity is easy to be lost.The similar action frames in a long time range are easy to lose.For the above problems,the reward mechanism was established in RDRL-GCN.In addition,aiming at the problems of large variance of frame selection network and difficult convergence,this paper introduces the reward mean into the objective function to reduce variance.To test the comprehensive performance of the algorithm,this paper compares RDRL-GCN with other cutting-edge algorithms in model parameters,GFLOPs and algorithm accuracy,and veri fies the feasibility and effectiveness of RDRL-GCN.