Research On The Method Of Action And Behavior Recognition In Rehabilitation Training

With the acceleration of social aging,the scale of chronic disease population is more and more large.Stroke as a typical chronic disease with the leading cause of death in China,and it has high incidence and disability rate.Rehabilitation exercise can restore most of the motor ability of stroke patients,and it is also the main treatment method of the patients in the rehabilitation period.However,the limited medical resources in our country are difficult to meet the rehabilitation needs of stroke patients at this stage.Long-term hospitalization will also bring expensive medical costs to the family.Therefore,most patients need to receive early treatment in the hospital and transfer to the family for follow-up rehabilitation training.In the home scene,it is difficult for the family members to do real-time supervision due to the reasons of going out to work and so on,which leads to the poor compliance of the patients in the family.Using intelligent monitoring technology can identify the rehabilitation actions of patients in real time from the video,so as to replace family members or even doctors to monitor and supervise the rehabilitation process of patients,which has a positive significance for the rehabilitation level of patients in the medium and long term.For the above requirements,this paper mainly researchs the action and behavior recognition problems in the rehabilitation training scene,especially the stroke rehabilitation actions.The main work is as follows:(1)A method of rehabilitation action recognition based on pose two-stream network is proposed.Firstly,the video frame is sampled from RGB video segment at medium intervals and the joints of human skeleton are extracted by openpose.There are missing joint points in the complex scenes,which are calculated from the skeleton data of adjacent frames by the Gauss weighted mean.Then the Holt smoothing method is used in the pose sequence to reduce jitter of joints coordinates,and the joint points are normalized based on the resolution of video.Then the spatial feature and motion feature are extracted from the pose sequence.The spatial stream network is designed based on the GRU and attention mechanism of multi-layer feature fusion,and the spatiotemporal information is extracted from the spatial feature sequence;the motion stream network is designed based on multi-layer 1D CNN and the combination of causal and dilated convolution,and the spatiotemporal information is extracted from the motion feature sequence.In the softmax layer,the adaptive weight fusion strategy is used to fuse the classification results of the two networks.The experimental results show that the two branches of the network achieved excellent recognition results,and the integrated network further improved recognition performance.The recognition rate in KTH reached 98.61%,and the recognition rate of 623 test videos in the rehabilitation action dataset reached 100%.(2)The samples of dataset are cut off-line video segments.In order to realize synchronous analysis of the real-time undivided video stream obtained by the surveillance camera,an online rehabilitation action recognition method based on the surveillance video stream is proposed on the basis of the pose two-stream network.There is interference from other irrelevant people in the family,and the pose sequence of the same person will lose the time sequence association in the multi-person scene.Firstly,an efficient single target tracking method is designed by combining the pose estimation and Kalman filter algorithm.The trajectory of the target body is tracked by predicting the position of target in next frame,and the corresponding online pose stream is generated at the same time.The sliding window is designed to intercept the segmented skeleton features from the pose stream.After preprocessing and feature extraction,the spatial stream network is used to recognize the segmented features.Then,the non maximum suppression algorithm is introduced to select the optimal window in the neighborhood,and the size is modified by multi-scale sub-windows.The experimental results show that the recognition rate of the model for online rehabilitation action is up to 91.8%,and the running speed on GTX 1060 is up to 18.30 FPS,which has a strong advantages in real-time and recognition rate.