基于Swin-TCN融合网络的运动视频理解的研究  

Research on Motion Video Understan ding Based on Swin-TCN Fusion Network

作　　者：魏士磊 付江龙[1] 王剑雄[1] 沈英杰 WEI Shilei;FU Jiangong;WANG Jianxiong;SHEN Yingjie(Hebei University of Architecture,Zhangjiakou 075000,Hebei)

机构地区：[1]河北建筑工程学院,河北张家口075000

出　　处：《长江信息通信》2024年第9期6-9,共4页Changjiang Information & Communications

基　　金：河北省体育科技研究课题资助项目(2024QT01);校级创新基金项目(XY2023080)。

摘　　要：针对运动比赛中运动员行为识别的问题,提出一种基于Swin Transformer和TCN(时域卷积网络)的混合网络模型。首先对视频进行预处理得到视频帧,使用Swin模块提取视频帧中人体动作相关的空间特征,将输出的特征图的空间尺寸压缩,纵向拼接后交给TCN模块提取视频中人体动作的时序特征。为提高通道对于行为识别结果的贡献度,在TCN残差块中加入通道注意力机制,经过分类模块后得到最终结果。实验结果显示,该模型在UCF101上动作识别准确率达到了89.7%。A hybrid network model based on Swin Transformer and TCN is proposed to address the issue of athlete behavior recognition in sports competitions.Firstly,preprocess the video to obtain the video frame,use the Swin module to extract spatial features related to human actions in the vidco frame,compress the spatial size of the output feature map,vertically concatenate it,and hand it over to the TCN module to extract temporal features of human actions in the video.To improve the contribution of channels to behavior recognition results,a channel attention mechanism is added to the TCN residual block,and the final result is obtained aftcr passing through the classification module.The experimental results show that the model achieves an action recognition accuracy of 89.7% on UCF101.

关 键 词：Swin Transformer 时域卷积网络 混合网络架构 行为识别 时序特征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]