Design Of Human Action Recogntion System For Rehabilitation Training Scene

In current society,with the rapid increase in the number of paralyzed and disabled people,patients urgently need to receive corresponding rehabilitation treatment in the process of physical recovery.At present,artificial rehabilitation and machine-assisted rehabilitation are mainly used in the field of rehabilitation treatment.However,manual rehabilitation will cost a lot of manpower and there are fewer professionals who specialize in it.At the same time,the high cost of rehabilitation robots is unacceptable to patients.Therefore,these traditional rehabilitation methods cannot meet the needs of society.With the rapid development of computer vision,action recognition technology based on deep learning is widely used in rehabilitation training due to its low cost and high efficiency.Based on this research background,this paper proposes a skeleton-based action recognition model based on graph convolutional neural network,conducts related experimental researches based on the model,and constructs a human action recognition system for rehabilitation training scenes.Under the guidance of the rehabilitation coaches and the supervision of the attending doctors,patients can effectively carry out the daily work of rehabilitation training,and the system evaluates the similarity and feeds back the results to the attending doctor,so as to monitor the process of the patients’ rehabilitation training.The specific research content is as follows:(1)Aiming at the problem that the existing graph convolution network model can not fully mine the spatio-temporal features,which leads to the poor recognition accuracy.Firstly,the graph attention mechanism in the spatial domain is used to allocate different weight coefficients according to the importance of neighbor nodes to generate an attention coefficient matrix,which can fully extract the spatial structure features of human body,furthermore,a new spatial self-adaptive adjacency matrix is proposed to furtherly enhance the extraction of spatial strcture features of human body combined with the global adjacency matrix generated by the non-local network.Then,mixed pooling model is utilized in temporal domain to extract key action features and global contextual features,these two-above features can be furtherly combined with the features generated by the temporal convolution to enhance the extraction of temporal features from behavioral informations;Furthermore,the ECA-Net network is introduced into the model to enhance the channel attention,which is beneficial to the model to extract the spatio-temporal characteristics of the sample.At the same time,combining the spatial feature enhanced,the temporal feature enhanced with the channel attention,an novel model referred to as STFE-GCN is constructed and one end-to-end training can be realized based on mutil-stream network to achieve the full mining of spatio-temporal features;Finally,in order to verify the effectiveness of the model in fully mining spatio-temporal features,the researches on skeleton-based action recognition are carried out on NTU-RGB+D and NTU-RGB+D120 datasets.(2)In order to effectively establish spatio-temporal context dependency,spatio-temporal and channel dependency,and strengthen the feature extraction of multi-level receptive field.Firstly,a spatio-temporal sampling graph convolution network is proposed,which takes sequential multiple frames as spatio-temporal sampling,and constructs local and global spatio-temporal context dependency by constructing spatio-temporal adjacency matrix to participate in graph convolution;Secondly,in order to effectively establish the dependency between spatio-temporal and channel and enhance the multi-level receptive field to capture more discriminative features,the temporal self-calibration convolution is proposed to convolution and feature fusion in two different space-time size : one is the space-time with original scale size,the other is the potential space-time with lower scale size;Then,by combining the spatio-temporal sampling graph convolution and the temporal self-calibration mechanism,the action recognition model is developed;Finally,the parameters optimization and the performance trials of the action recognition model are accomplished on large public datasets.(3)In order to meet the needs of patients’ daily rehabilitation training in the home environment,an action recognition system for rehabilitation training scenes was designed and constructed.First of all,through the analysis of rehabilitation training actions and the datas collection by Kinect camera,an action dataset for rehabilitation training scenes is developed;Secondly,the DTW-LCS algorithm,which combines the DTW algorithm and the LCS algorithm,is proposed to be applied to the system as a motion evaluation algorithm to solve the problem of evaluating the effect of rehabilitation training for patients;Then,by selecting some actions in the action recognition dataset for rehabilitation training scenes,the experiments verify the high recognition accuracy of the system and the feasibility of the DTW-LCS algorithm;Finally,bying combining with the hardware selection and the software function requirements,an action recognition system for rehabilitation training scenes is designed and the relevant operation interfaces are displayed.