复杂运动场景下的多人姿态估计研究  

Research on Multiplayer Pose Estimation in Complex Sports Scenes

作　　者：柳长源 臧彦丞 兰朝凤 LIU Changyuan;ZANG Yancheng;LAN Chaofeng(College of Measurement and Control Technology and Communication Engineering,Harbin University of Science and Technology,Harbin 150080,China)

机构地区：[1]哈尔滨理工大学测控技术与通信工程学院,哈尔滨150080

出　　处：《电子科技大学学报》2024年第6期930-939,共10页Journal of University of Electronic Science and Technology of China

基　　金：国家自然科学基金(11804068);黑龙江省交通运输厅科技项目(HJK2024B002)。

摘　　要：针对运动场景中运动员之间的相互遮挡、自身部位遮挡、运动器械遮挡及复杂背景干扰问题,提出一种高分辨特征生成复原网络,引入融合注意力机制筛选有用特征信息通道,加入反卷积和多尺度特征融合模块分层处理小目标人像与大中型目标人像的姿态估计任务。设计生成对抗模块,对缺失部分进行补全和预测得到关节点热图,经过姿态骨架和最优匹配算法确定出关节点连接方式,并输出可视化姿态估计结果。在MSCOCO和Crowd Pose数据集上的实验结果表明该姿态估计方法在复杂运动场景下效果更优。In terms of the problems such as mutual occlusion,self-occlusion,sports equipment occlusion and complex background interference among athletes in motion scenes,a high-resolution feature generation recovery network is proposed in this paper.The attention fusion mechanism is introduced to screen the useful feature information channels.The deconvolution and multi-scale feature fusion modules are added to deal with the pose estimation tasks for small target portraits and large and medium-sized target portraits in a hierarchical manner.The adversarial module is designed and generated to complete and predict the missing parts to obtain the keypoint heat map,the keypoint connection mode is determined through the pose skeleton and the optimal matching algorithm,and the visual pose estimation results are output.Experimental results on MSCOCO and Crowd Pose datasets have showed that the pose estimation method is more effective in complex motion scenes.

关 键 词：人体姿态估计 深度学习 复杂运动场景 融合注意力机制 生成对抗网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]