Rehabilitation Training Action Correction System Based On Human Pose Recognition

Health is the common pursuit of all mankind and the cornerstone of national prosperity.In the field of sports medicine,scientific and effective rehabilitation training can help patients regain their motor function as soon as possible.At the moment when the supply of nursing staff is in short supply,it is difficult to fully popularize professional guidance in the face of more demand for rehabilitation training.When most patients perform self-training,the training movements are not standardized,resulting in low rehabilitation efficiency and even damage to their own bodies.Traditional solutions usually require additional wearing of sensor equipment,such as exoskeleton robots.Although they can help patients accurately identify movement specifications,they are basically only suitable for professional medical institutions due to inconvenient wearing and high costs.In order to solve the above problems and enable most patients to obtain standardized training guidance,this paper explores how to use artificial intelligence technology to build a low-cost,high-performance rehabilitation training movement correction system.The main work is as follows:1)Based on the two gesture recognition algorithms of MoveNet and OpenPose,an improved gesture recognition algorithm model is proposed,which realizes real-time and accurate gesture recognition.These include improving the feature extraction network VGGNet of OpenPose for the needs of rehabilitation training scenarios,optimizing it and introducing it into the lightweight gesture recognition network MoveNet.These include the full convolution processing of this part of the feature extraction structure,the introduction of residual items,and the input and output adaptation of the back-end feature prediction and classification network.The experimental results show that the improvement can effectively improve the prediction accuracy of the model.In the test on mobile devices,the processing speed of the improved pose recognition model under the enhanced COCO dataset can reach 39 frames per second,with an average accuracy rate of 82.7%.2)Based on the principle of space coordinate system transformation and the fully connected neural network structure,a motion correction algorithm is proposed,which realizes real-time and accurate motion correction.This includes data set generation,network training optimization,etc.When the data set is generated,in order to simulate various training postures of the patient under different camera angles,the skeleton data information of the standard action after spatial transformation is randomly transformed several times to generate an enhanced training action and correction vector data set.On this basis,the corresponding relationship between the fitting training action and the correction vector is realized through a lightweight fully connected neural network,and the influence of various irrelevant factors such as camera angle deviation,lens distortion,and body characteristics on the action correction results is eliminated.The experimental results show that the algorithm has a good effect on correcting wrong actions.In the mobile device test,the processing speed reaches 35 frames per second,and the correction error is 3.7%.3)Based on the above work content,a rehabilitation training action correction system based on gesture recognition is designed and implemented.The system can be built on a PC or mobile device.First,the camera is used to obtain the patient’s training actions,the skeleton coordinate information is output through the gesture recognition model,and time domain filtering is performed to smooth the temporal features,and then the real-time and accurate correction vector is output through the action correction network,and finally realize the visualization of the correction results,forming a complete set of posture correction system structure.