摘要
人体姿态估计是计算机视觉领域的研究热点之一,目标是从给定的图像或视频中识别人体的关节.由于篮球比赛视频中人体动作复杂多变,易产生运动模糊、遮挡等问题,导致现有的人体姿态算法对篮球动作姿态估计的准确率较低.针对这一问题,提出了一种基于多尺度时空关联特征的篮球动作姿态估计算法,构建基于Transformer的人体时序特征捕捉模块对序列层级的时空特征信息进行建模,以缓解运动模糊、遮挡等现象带来的负面影响.此外,针对人体外形复杂多变的问题,提出了基于可形变卷积的人体空间特征残差融合模块来获取更为充分的空间特征.与现有算法相比,该算法在自行构建的篮球场景人体运动数据集、姿态估计公开基准数据集PoseTrack2017和PoseTrack2018均取得较好的效果.
Human pose estimation is one of the research hotspots in computer vision,which aims to locate the anatomical keypoints of all persons in an image or a video.Generally,the human movements in basketball videos are characterized by complexity and variability,which may cause motion blur,and severe occlusion.Therefore,existing human pose estimation algorithms yield low accuracy for pose estimation of basketball actions.To tackle this problem,a basketball action pose estimation algorithm based on multi-scale temporal association features is proposed,which alleviates the problems of motion ambiguity and severe occlusion by constructing a Transformer-based human body temporal feature capture module that models sequence-level temporal features.Furthermore,to adapt the complex and variable human body shapes,a deformable convolution-based human spatial feature residual fusion module is proposed to acquire more comprehensive spatial features.Compared with existing algorithms,the proposed method achieves better results in the self-built human motion dataset for basketball scenes and the pose estimation benchmark datasets PoseTrack2017 and PoseTrack2018.
作者
马骁
闫育东
MA Xiao;YAN Yudong(College of Physical Education and Science,TianJin University of Sport,Tian Jin 300000,China;Sports Department,Shenyang Aerospace University,Shenyang 110136,China;College of Continuing Education,TianJin University of Sport,Tian Jin 300000,China)
出处
《中南民族大学学报（自然科学版）》
CAS
北大核心
2023年第1期95-102,共8页
Journal of South-Central University for Nationalities：Natural Science Edition
基金
天津市研究生科研创新资助项目(2019YJSB096)。
关键词
姿态估计
时空建模
卷积神经网络
视频序列建模
pose estimation
spatio-temporal modelling
convolutional neural networks
video sequence modelling
