Research On Athlete Action Recognition Method Based On Deep Learning

With the rapid development of artificial intelligence,human action recognition technology has received high attention and is widely used in various fields such as smart home,medical rehabilitation,intelligent security,and motion analysis.Traditional action recognition methods rely on manual feature extraction,which is difficult to use in the case of large data.The action recognition method based on convolutional neural network can automatically extract features,which greatly simplifies the process of manual feature extraction in traditional methods,but it is easily disturbed by the background environment.In view of the above situation,this thesis proposes the S-ET-GCN action recognition method based on HRNet-DSC-CBAM bone key point detection,which can better describe action features by introducing bone sequence information.The main research work of this thesis is as follows:Firstly,ten types of Sanda actions are collected for the construction of the experimental data set.In order to improve the recognition efficiency and accuracy of the network,the method based on video clustering is used for key frame extraction and the method of bilateral filtering for image denoising.Secondly,in order to enhance the effective features of the input data and suppress its invalid features,the Yolov5s-CBAM target detection network model is obtained by fusing CBAM module and Yolov5 s algorithm,which improves the detection accuracy of human targets;Aiming at the problem that the increase in the number of network layers of the HRNet algorithm will increase the amount of calculations and parameters of the residual module,the HRNet-DSC-CBAM bone key point detection network model is designed,and the network model adopts a light residual module composed of DSC convolution and CBAM module fusion to replaces part of the original residual module,which reduces the amount of calculation and parameters and improves the accuracy of key point detection of human bones.In addition,in order not to affect the subsequent feature extraction,the continuity between actions is used to fill the bone key points lost due to occlusion,and the detected human bone key points are optimized.Finally,aiming at the problem that the temporal convolutional network in the ST-GCN algorithm cannot fully extract temporal features,the S-ET-GCN action recognition network model is designed.Based on the ST-GCN algorithm,the network model extends the TCN network layer and adds a residual mechanism,which better captures the short-term and long-term time dependencies between actions.At the same time,the DTW algorithm is used to calculate the joint angle difference between the standard action sequence and the test action sequence,and the action evaluation formula is defined to evaluate the Sanda action.The experimental results show that the target detection network model proposed in this thesis can more accurately locate the position of the human body in the image;the human skeleton key point detection network model can detect the key points of the human skeleton more quickly and accurately based on the obtained human body position;the action recognition network model can more accurately identify Sanda actions based on the obtained skeleton data.The recognition accuracy rate on the self-made Sanda dataset has reached 94.8%,which is 2% and 6.5% higher than that of the basic ST-GCN algorithm and LSTM algorithm respectively.At the same time,the action evaluation formula can realize the effective evaluation of the Sanda action and realize the fair and just auxiliary scoring.