Table Tennis Based On Improved GoogLeNet Network Human Motion Recognition

With the increasing maturity of deep learning technology,human action recognition based on deep learning has received widespread attention from domestic and foreign research scholars.Technical action detection in sports videos is an important application of computer vision in sports.Among them,table tennis has distinct technical characteristics,which has significant advantages for the recognition and classification of single human action technology.Therefore,the recognition of human movement in table tennis based on deep learning has important research significance and application value.This thesis identifies table tennis action on the basis of GoogLeNet network.Based on the understanding and analysis of Inception V1,Inception V2,Inception V3 network structure,according to the characteristics of human movement in table tennis,the middle and lower layers of GoogLeNet network are improved from the Inception module.Select Leaky-Relu activation function and Dorpout policy and Adam optimizer in network training policy.In the establishment of the data set,the table tennis action video in UCF101 and self-made data video are used,and the image data set is established after frame cutting.The expanded data set is established for the image data set through image flipping and mirror image,image blurring and random brightness change.The improved GoogLeNet network table tennis action recognition code is completed on Tensorflow,and the table tennis action recognition effects under different Dropout ratios,different optimization method combinations,different Adam optimization parameters and different network types are verified on the data set.The identification results show the effectiveness of the improved network.In view of the problem that color images images cannot reflect the continuity of human movements in table tennis video and that single-flow network cannot effectively extract movement information in time dimension,the improved GoogLeNet is used as the basic network framework to build a spatio-temporal dual-flow convolutional neural network for table tennis action recognition.The improved GoogLeNet network can effectively extract the static information of spatial flow RGB images and the motion information of time flow stacked optical flow images,and the extracted features are fused by weighted fusion method in the later classification layer.At the same time,after feature fusion,in order to classify technical movements of table tennis more accurately,this thesis applies the am-Softmax classification algorithm improved based on Softmax classifier to human movement recognition of table tennis.The experimental results show that the table tennis action recognition method based on the feature fusion of improved GoogLeNet dual-flow convolutional neural network and the introduction of AM-Softmax classification algorithm can improve the recognition ability of the model and the training speed of the model.