运动场景下的多目标人体姿态估计  

Multi-target human pose estimation in sports scenes

作　　者：朱丽萍[1] 唐亮 朱凯杰 李昊阳 ZHU Li-ping;TANG Liang;ZHU Kai-jie;LI Hao-yang(College of Information Science and Engineering,China University of Petroleum(Beijing),Beijing 102249,China;College of Sports,Humanities and Arts,China University of Petroleum(Beijing),Beijing 102249,China;Faculty of Science and Technology,University of Macao,Macao 999078,China)

机构地区：[1]中国石油大学(北京)信息科学与工程学院,北京102249 [2]中国石油大学(北京)体育与人文艺术学院,北京102249 [3]澳门大学科技学院,中国澳门999078

出　　处：《计算机工程与设计》2023年第7期2156-2162,共7页Computer Engineering and Design

基　　金：国家重点研发计划基金项目(2019YFA0708300);国家自然科学基金项目(52074323)。

摘　　要：为提高群体活动场景下细粒度人体姿态估计的准确率,优化网路中人体识别及姿态估计算法,在现有研究的基础上,提出一种结合多尺度预测以及改进并行注意力模块的多目标人体姿态估计算法。在充分利用不同尺度特征信息的基础上,实现高质量的人体姿态估计;针对运动场景下多目标人体姿态数据集较少,提出一种数据集CUPB Sport Dataset。实验结果表明,该算法在公开基准数据集和自制数据集上分别达到了81.4 mAP和79.7 mAP,验证了该算法在运动场景下针对多目标的高效性。To improve the accuracy of fine-grained human pose estimation in group activity scenarios,and optimize the human body recognition and pose estimation algorithms in the network,based on the existing research,a multi-scale human pose estimation method combining multi-scale prediction and improved parallel attention module was proposed.On the basis of making full use of feature information of different scales,high-quality human pose estimation was achieved.A dataset CUPB Sport Dataset was proposed for the lack of multi-target human pose datasets in sports scenes.Experimental results show that the algorithm achieves 81.4 mAP and 79.7 mAP on public benchmark datasets and self-made datasets,respectively.It is verified that the algorithm is efficient for multi-targets in moving scenes.

关 键 词：深度学习 运动场景 姿态估计 注意力机制 多目标 多尺度 细粒度人体识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]