基于长短期记忆-卷积神经网络(LSTM-CNN)的体育训练视频中错误动作的识别方法  

A Method for Identifying Erroneous Actions in Sports Training Videos Based on Long Short Term Memory-Convolutional Neural Network(LSTM-CNN)

作　　者：徐静 何敬堂 XU Jing;HE Jingtang(College of Physical Education,Huainan Normal University,Huainan Anhui 232001,China)

机构地区：[1]淮南师范学院体育学院,安徽淮南232001

出　　处：《上饶师范学院学报》2024年第6期81-91,共11页Journal of Shangrao Normal University

基　　金：安徽省高等学校人文社会科学研究重点项目(SK2023A0507);淮南师范学院校级横向课题(2024HX051)。

摘　　要：为了提取丰富的高级空间特征、减少关键信息的损失、全面捕捉视频中的时序特征、提升视频中错误动作的识别效果,研究了一种基于长短期记忆-卷积神经网络(long short term memory-convolutional neural network,LSTM-CNN)的体育训练视频中错误动作的识别方法。第一步,通过第一层卷积神经网络(convolutional neural network,CNN)层提取体育训练视频中错误动作的低级空间特征;第二步,利用第二层CNN层在体育训练视频低级空间特征内提取其高级空间特征;第三步,通过两层CNN层逐步提取体育训练视频的空间特征,确保在提取其高级特征的同时尽量减少其关键信息的损失,保证提取的高级空间特征具有丰富性;第四步,利用长短期记忆网络(long short term memory,LSTM)网络层,在高级空间特征内,对体育训练视频中错误动作的时序特征进行全面提取;第五步,引入注意力机制,对体育训练视频中错误动作的时序特征进行筛选,获得更有价值的时序特征,进一步提升错误动作的识别效果;第六步,通过归一化指数(Softmax)分类器,结合筛选出来的时序特征,输出层将体育训练视频中错误动作的识别结果输出。实验证明,基于LSTM-CNN的体育训练视频中错误动作的识别方法可有效提取体育训练视频动作的空间特征,并可在不同场景下精准识别体育训练视频中的错误动作。In order to extract rich advanced spatial features,reduce the loss of key information,comprehensively capture temporal features in videos,and improve the recognition effect of erroneous actions in videos,a method for identifying erroneous actions in sports training videos based on long short term memory-convolutional neural network(LSTM-CNN)was studied.The first step was to extract low-level spatial features of erroneous actions in sports training videos through the first convolutional neural network(CNN)layer;The second step was to use the second CNN layer to extract high-level spatial features from the low-level spatial features of sports training videos;The third step was to gradually extract the spatial features of sports training videos through two layers of CNN,ensuring that the loss of key information was minimized while extracting their advanced features,and ensuring that the extracted advanced spatial features were rich;The fourth step was to use a long short term memory(LSTM)network layer to comprehensively extract timing features of erroneous actions in sports training videos within advanced spatial features;The fifth step was to introduce attention mechanism to screen the timing features of erroneous actions in sports training videos,obtain more valuable temporal features,and further improve the recognition effect of erroneous actions;The sixth step was to use a normalized index(Softmax)classifier,combine with the selected timing features,and output the recognition results of erroneous actions in the sports training video by the output layer.Experimental results have shown that the recognition method of erroneous actions in sports training videos based on LSTM-CNN could effectively extract spatial features of actions in sports training videos and accurately identify erroneous actions in different scenarios.

关 键 词：长短期记忆网络 卷积神经网络 体育训练视频 错误动作识别 空间特征 时序特征 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]