Design And Implementation Of Table Tennis Action Classification And Comparison System Based On Machine Vision

With the rapid development of machine vision technology in recent years,the rapid increase of various video data has made behavior recognition based on video data become one of the hot research topics.The table tennis action classification and comparison system based on machine vision designed in this paper focuses on the application of deep learning-based motion recognition technology to intelligent table tennis training scenarios.In the absence of a coach's judgment,the system only collects user training action videos to identify actions and analyzes the completion of athletes' technical action training.The system designed in this article adopts the design architecture of C/S.The client program data acquisition module acquires and displays the real-time video stream and transmits it to the server synchronously.After the server receives the data,it performs motion recognition and feedbacks the recognition result.The client and server use multi-threaded technology for data synchronization.Aiming at the shortcomings of traditional two-stream convolutional network in processing long-term information,the sparse sampling-based time segmentation two-stream network adopted in this paper can better express the long-term motion features.First,the continuous video frame data is divided into multiple segments,and each segment of video frame sequence is randomly sampled to form a short sequence of data containing user actions,and then it is applied to the two-stream network for feature extraction.The extraction of optical flow images involved in the two-stream network is implemented by the Lucas-Kanada algorithm.The optical flow contains the movement time information of the target,which can effectively represent the movement of pixels in different areas of continuous frame images.At the same time,due to the small amount of action video data,the system adopted a variety of data enhancement processing and network pre-training to mitigate the risk of overfitting during network training.Finally,this paper analyzes the feature fusion methods of multiple two-stream networks,and proposes to adopt the network fusion method in the convolution layer and use 3D convolution and pooling operations to perform feature aggregation operations,so that the network can more effectively express the action of spatio-temporal characteristics ensure high recognition accuracy.The system has been tested in actual scenarios,and the results show that the system design meets the expected requirements.