Research And System Implementation Of Ping-pong Motion Recognition Based On Bone Information

With the progress of spiritual civilization and the rapid development of material civilization in human society,people’s way of life has undergone dramatic changes.However,they are facing serious personal health problems while enjoying efficient and convenient life.The lack of sports in modern people is one of the important reasons affecting health,and using effective methods to exercise and promote physical health will become an important part of modern life.Table tennis as a national sport in China is suitable for many enthusiasts and adjustable amount of exercise.However,the traditional table tennis is limited by time and field,and long-term incorrect movement posture will cause physical injury.At present,computer vision-based human motion recognition technology is a hot research topic,but the related work is mainly reflected in visual interaction,there is no recognition and evaluation research on table tennis motion.Also,it is difficult to recognize human ping-pong action in real time since its complexity is high,and the movement process is easily blocked.Based on computer vision technology,this paper used bone key detection method to obtain key point information of human bones,built a classification model to identify typical movements of ping-pong,and used DTW algorithm to evaluate human table tennis actions in real time.Finally,the above parts were applied to human table tennis action recognition and evaluation system,which provided table tennis lovers with skills improvement free of time and space constraints.At the same time,it provided a reference for the study and improvement of table tennis movement,made somatosensory movement more scientific to promote health,and provided an application example for other sports.The main work is as follows:1.Research on bone point algorithm and extraction of bone key points.Firstly,the selection of bone key detection method was introduced,and three classical bone key detection models,Open Pose,Alpha Pose and Lightweight Open Pose,were tested and compared,and Light Weight Open Pose was selected as the bone key detection method in this paper.Secondly,the network architecture of Open Pose model was introduced in detail,the training process of the two branches of neural network was described,the calculation method of confidence and affinity was explained,and the method and principle of bone key detection were expounded.Thirdly,the improved network architecture of Lightweight Open Pose model was introduced,and the optimization part of the network backbone,network design,and refinement phase in Light Weight Open Pose were described.Finally,through the lightweight Open Pose skeleton key point detection model,the human table tennis action data set video and real-time action video were detected.2.Human table tennis action data set production,data preprocessing and feature extraction.First of all,the human table tennis action data set was made.Ten college students’ human table tennis actions were collected via the camera.The video frames of the collected continuous human table tennis actions were separated and reasonably clipped by the three frame difference method,and then input into the lightweight Open Pose model.The key points of the skeleton extracted from the video were extracted into the data set in CSV format.At the same time,to reduce the influence of irrelevant factors on the recognition accuracy,the data set was pre-processed by deleting redundant key points,adding missing key points and filtering.Finally,feature extraction was carried out.The neck reference method,posture-to-angle conversion method and coordinate unification method were compared,and the coordinate unification method was chosen as the feature extraction method of this paper.3.Research on the construction and evaluation method of human table tennis action recognition model.Firstly,the human table tennis action recognition model was constructed,the classification model was trained and tested and the control experiment was set up by K-neighbor,support vector machine and long-term and short-term memory network.The experimental results showed that the support vector machine classification model exhibited high accuracy in real-time recognition of human table tennis action in self-made data set.Therefore,this paper used support vector machine to build human table tennis action recognition model.Secondly,the motion evaluation method was studied.The DTW algorithm was used to calculate the cumulative distance between the real-time human table tennis action and each human bone key point in the standard human table tennis action sequence.The similarity between the bone key points was calculated through the cumulative distance,and the human bone key points were scored and the low score bone key points were marked.The experimental results revealed that this paper achieved a better understanding of human table tennis action real-time scientific assessment.4.Design and implementation of pingpong health promotion system.Based on the above research content,we designed and implemented the human table tennis action recognition and evaluation system.After detailed demand analysis,we divided the system into data training module,action recognition module and action evaluation module,and introduced the three modules in detail.Finally,we completed the implementation of the system and displayed the function and interface.