Recognition Of Basic Table Tennis Techniques Based On Spatial Temporal Graph Convolutional

Table tennis has a special status in China,and table tennis players have won countless honors for the motherland on the international arena.This is not only related to the talents and efforts of the athletes themselves,but also closely related to the efforts of the coaching team behind them.The coaching team relies on "table tennis technical and tactical analysis" to continue targeted training.The traditional method relies on manually marking the ball hitting action,which takes time and effort.In recent years,with the popularity of deep learning,it has been widely used in the field of daily motion recognition.The classification of batting movements through deep learning can improve training efficiency,but unlike daily movements,table tennis has additional movements and occlusion problems.At the same time,there is currently no data set for the classification of table tennis batting movements.In response to the above problems,this article first established a video-based basic table tennis action data set.The data set covers seven basic technical actions in table tennis competition,namely: forehand attack,backhand attack,forehand chop,backhand chop,forehand push,backhand push and block.Combining with the existing research,this paper first uses the relevant methods of human body posture prediction to preprocess the video data of basic table tennis technical movements,and converts the video data into bone data.For these skeletal data,in the third chapter of this paper,we propose a motion recognition method based on local joint points in table tennis basic technology under the convolution framework of spatiotemporal graphs.Specifically,the training network for human posture prediction is modified to output local three-dimensional coordinates of nine joint points.The three-dimensional coordinate information of these 9 joint points is input into the convolution network of the space-time graph,and the network parameters are trained to learn the spatial and temporal characteristics of the joint points.Finally,the trained parameters are used to test and output the accuracy of action classification.This method can achieve a recognition rate of 93.76% on the self-built basic table tennis action data set,and a recognition rate of 79.41% on the public data set NTU RGB+D.At the same time,training directly using video data can also better learn the contour information and color information of the video.Therefore,in the fourth chapter,this paper proposes a dual-stream fusion method combining video data and bone data.In this method,the video data is manually cropped,and the cropped video image is input to the convolutional neural network frame by frame.By adding an attention mechanism,the network spontaneously pays attention to the athlete's hand movements,and the data with the attention mechanism is input into the convolution long-term memory Network training and testing.This method can achieve a recognition rate of88.89% on the self-built table tennis basic technical action data set.At the same time,the average value of the results of this method and the results of the method in Chapter3 is output as the final result,which can achieve a variety of information fusion.Finally,the self-built table tennis basic technical action data set classification recognition rate was increased to 94.97%.