| With the rapid development of society and economy,people’s living standards are improving day by day,and modern people are paying more and more attention to their own health problems and want to achieve the effect of physical fitness through various sports.Among them,ball sports(badminton,tennis,table tennis,etc.)are popular among fitness enthusiasts.At the same time,with the breakthrough of Kinect in the human bone tracking algorithm and the rise of deep learning,the research of human posture recognition based on joint point data and deep learning methods has gradually become a new research trend.At present,the gesture recognition of ball sports is very challenging in terms of movement complexity and research data.At the same time,the research also has extremely important application value.Therefore,this thesis takes badminton as a representative,combined with the analysis and summary of the existing human body gesture recognition research,based on the joint point sequence and LSTM neural network,to carry out the research on the human body ball motion gesture recognition problem.The research content of this article mainly includes:(1)Aiming at the problem of recognition errors or loss when Kinect collects human bone joint point data,a joint point information processing method based on bone length and motion continuity is proposed.This method can evaluate the credibility of the joint point information,and repair the detected data with errors.By repairing the error data,more effective joint point information can be obtained.Experiments show that this method can effectively improve the accuracy of ball movement gesture recognition.(2)In order to fully obtain the feature information related to the ball movement posture,a ball movement posture recognition method based on the LSTM-Attention network is proposed.This method first performs feature enhancement preprocessing based on angle and relative distance on the joint point information,so as to enrich the feature information and improve the convergence speed of the network.Secondly,combining LSTM and attention mechanism to recognize and classify ball movement postures,so as to further mine the deeper feature information in the sequence.At the same time,by introducing dropout in the network to reduce network over-fitting.This method not only improves the efficiency of network training,but also effectively improves the accuracy of human body ball movement gesture recognition.(3)Designed and developed a Kinect-based ball movement gesture recognition system,which mainly includes three modules of ball movement data collection,processing and analysis.The system can automatically collect the joint point data of the ball movement posture,and process the acquired data,and then the ball movement posture recognition network outputs the user’s posture prediction result.Based on the results,the system will give a corresponding action comparison chart for analysis,so as to help users improve the standardization of ball sports postures. |
