Research On Human Behavior Recognition Method Based On LSTM

Human motion recognition has become an active area of computer vision,and there are many important research issues,such as event recognition,group-based activity recognition,human-machine interaction,and human activity analysis in video.Most of the methods currently proposed are to recognize the actions in the RGB video recorded by the 2D camera.However,it is still a challenging issue for three reasons.Firstly,it is difficult to extract useful information from high-dimensional and low-quality input data.Secondly,RGB video is very sensitive to certain factors,such as illumination changes,occlusion,and background clutter.Thirdly,the recognition of motions and such as human poses and objects.Related to advanced visual cues of classes,these clues are difficult to obtain directly from RGB video.Humans can identify movements by describing the main joint points of the bones in the behavior,and experiments have shown that a large number of behavior categories can only be identified from the bones.Compared to RGB-based motion recognition,skeleton-based motion recognition avoids the horrible task of extracting features from video and can explicitly simulate the dynamics of motion.There are three ways to get the skeleton:motion capture system,RGB image and depth map.Complex motion capture systems are very expensive and require users to wear a motion capture kit with markers;extracting a reliable skeleton from a single RGB image or video,ie pose estimation,is still a very difficult problem.At present,the existing algorithms mostly construct the network model from the perspective of time and space of the skeleton sequence,and do not consider the construction of the network model from the relationship inside the skeleton sequence.Therefore,this paper proposed the TLSTM-Atten neural network model to extract the internal dependence of the skeleton sequence.Previously,the Self-Attention mechanism was studied and analyzed in detail,and finally the Softmax function was used to perform the final recognition rate statistics.In this paper,the proposed TLSTM-Atten neural network is applied to the NTU RGB+D and Northwestern-UCLA data sets,and the experimental results are analyzed in detail.On the NTU RGB+D dataset,this paper uses the TLSTM-Atten neural network to train in the original samples.The recognition rate is increased by nearly 3%on the CS and CV standards respectively.On the Northwestern-UCLA dataset,this paper uses TLSTM-The Atten neural network trained in the original sample reached 89.49%,which was 15.29%better than the Lie Group network,10.83%better than the HBRNN-L network,and 3.50%better than the TS-LSTM network.In summary,the main contributions of this article are as follows.First,the Kinect 2.0 sensor is calibrated to analyze the principle of bone data extraction.Secondly,a TLSTM-Atten neural network is proposed to extract the internal dependence between the skeleton sequences.Third,the TLSTM-Atten neural network application The NTU RGB+D dataset and the Northwestern-UCLA dataset were demonstrated and demonstrated the validity of the TLSTM-Atten neural network for recognition based on skeleton-based behavior.Finally,the experimental results are analyzed from the perspectives of recognition rate,confusion matrix and histogram.