Research And Implementation Of Ping-pong Action Recognition System Based On Wearable Inertial Sensing Data

Human action recognition is a research hotspot in the field of pattern recognition.It processes and analyzes the data collected by sensors through a computer,learns to understand human movements and behaviors,and makes corresponding decisions.It is increasingly used in sports and other fields,giving users a personalized exercise evaluation program,thereby helping people improve their sports skills and enhance human health.In recent years,with the rapid development of wearable computing,human action recognition based on wearable inertial sensors has attracted a large number of researchers.Compared with visual sensors,wearable inertial sensors have many advantages such as low cost,small size,wider application range,no space restrictions and occlusion,better protection of user privacy and so on,which can enable users to obtain more free movement space.It is more suitable for use in sports.Table tennis is a popular ball sport that is loved and sought after by people from all over the world,especially the Chinese people.It enjoys a broad mass base and profound cultural heritage in China,promoting the integration of national fitness and health.At present,there are few studies on Ping-pong action recognition and evaluation systems,so this article uses wearable inertial sensors to construct a low-cost,low-latency,and highaccuracy human action recognition and evaluation system for complex Ping-pong action recognition and evaluation.The main research work of this paper is as follows:(1)Implement high-performance server applications based on the Socket AsyncEvent Args class.In this study,we used the asynchronous events,object pool technology,data buffer pool,reducing the frequent creation and destruction of runtime threads and other technologies provided by the Socket Async Event Args class to implement highperformance server applications,and solves the simple Socket to deal with multi-client node high concurrent data communication Insufficient capabilities,while designing and implementing a multi-sensor data synchronization strategy.Experiments have proved that high-performance server applications based on the Socket Async Event Args class can also obtain stable communication capabilities and low-latency data transmission effects on low-cost hardware devices.(2)Proposed window segmentation point detection and Key-frames extraction methods.Based on the "3σ" principle of the normal distribution,this paper defines Inertial Data Key-frames.A window segmentation point detection and key frame extraction method is proposed to obtain effective key frame data from the real-time inertial sensing data stream by window segmentation.And realized respectively through two threshold value discrimination algorithms of the mean and the difference,respectively.Experiments show that this method can more accurately extract the action Key-frames from the real-time inertial sensing data stream,remove a large amount of redundant data,and play an important role in the retrieval,analysis,and real-time recognition of motion actions.(3)Improved Inception network structure for multi-dimensional feature extraction.In this study,on the basis of the 2-dimensional convolution in the Inception network structure,a 1-dimensional convolution is added to extract the features of the inertial data in the time series,expand the dimension of the feature map and enhance the data expression ability and generalization ability.Experiments show that adding an improved Inception network before the convolutional neural network can significantly improve the classification accuracy of human Ping-pong actions,and significantly improve the recognition effect of similar Ping-pong actions.(4)Propose an evaluation method for human Ping-pong action.This study evaluates the similarity between the overall Ping-pong action and the partial action and the standard action.The probability value of the Ping-pong action classification result is used to evaluate the similarity between the overall action and the standard action.Human body local action evaluation uses the cosine similarity algorithm to calculate the similarity between each sensor feature and the standard feature vector as the local action evaluation result.Experiments show that the human Ping-pong action evaluation method proposed in this paper has certain accuracy and guiding significance for the evaluation results of Ping-pong action.(5)Realize a high-performance Ping-pong action cloud recognition and evaluation system.Based on the basic theory of software engineering,the feasibility analysis,requirements analysis,system architecture and main function summary design,system function realization and system deployment of high-performance Ping-pong action cloud recognition and evaluation system are completed,and a system with good robustness and stable operation can be realized.In summary,this paper studies data collection,window segmentation point detection and key frame extraction in the process of wearable inertial perception Pingpong action recognition,human Ping-pong action modeling,and human Ping-pong action evaluation methods.High-performance Ping-pong action cloud recognition and evaluation system.