摘要
针对足球赛事视频中动作识别存在的时序定位精度不足、类别分布不均衡等问题,提出了一种基于改进E2E-spot轻量级时空分离注意力算法(LSSA-Net)。在特征提取时,提出的轻量化时空分离注意力模块(LSSA)能够增加特征提取能力;在LSSA模块时间分支中设计因果卷积与可学习时序编码联合架构,提高算法时序定位精度;使用Logit Adjustment作为算法的损失函数,解决足球比赛中动作类别不平衡的问题。该算法在SoccerNet-v2数据集上与其它方法比较,结果显示在mAP(Tight和Loose值)上分别比E2E-spot提升了4.84%和2.09%,并且相比其它方法也有不同程度的提升。
To address the issues of insufficient temporal localization accuracy and imbalanced class distribution in action recognition within soccer match videos,a lightweight spatio-temporal separation attention algorithm(LSSA-Net)based on improved E2E-spot was proposed.During feature extraction,the feature extraction capability was enhanced by the proposed lightweight spatio-temporal separation attention(LSSA)module.Causal convolution and learnable temporal encoding joint architecture were designed in the time branch of the LSSA module to improve the temporal localization accuracy of the algorithm.Logit Adjustment was employed as the loss function to address class imbalance in soccer action recognition.Evaluated on the SoccerNet-v2 dataset,LSSA-Net outperforms existing methods with a 4.84%and 2.09%improvement in mAP(Tight and Loose values)over the baseline E2E-spot,while also demonstrating superior performance against other comparative approaches.
作者
魏志珍
黄佳旺
陈文文
刘城宇
陈艾东
WEI Zhi-zhen;HUANG Jia-wang;CHEN Wen-wen;LIU Cheng-yu;CHEN Ai-dong(Department of Physical Education,Beijing Union University,Beijing 100101,China;Beijing Key Laboratory of Information Service Engineering,Beijing Union University,Beijing 100101,China;College of Robotics,Beijing Union University,Beijing 100101,China;Research Centre for Multi-Intelligent Systems,Beijing Union University,Beijing 100101,China)
出处
《计算机工程与设计》
北大核心
2025年第8期2163-2169,共7页
Computer Engineering and Design
基金
国家重点研发计划基金项目(2022YFB2804402)。
关键词
足球赛事视频
定位精度不足
类别不平衡
轻量级时空分离注意力
特征提取
因果卷积
可学习时序编码
损失函数
soccer match videos
insufficient localization accuracy
class imbalance
lightweight spatio-temporal separable attention
feature extraction
causal convolution
learnable temporal encoding
loss function
