文章研究了基于视频的乒乓球球员动作识别问题。在计算机视觉领域，人体动作识别具有一定挑战性。基于专业乒乓球运动员在乒乓球发球机的接发动作视频，构建了乒乓球击球动作视频数据集，将其分为正手击球、反手击球、正手拉球、反手拉球和非击球动作5类。提出通过人体密集姿态(Dense Pose)处理数据集，将把人体形态从环境中进行提取，随后提出一种改进的C3D卷积网络，用于学习数据集上连续帧的时空特征。结果表明，文章设计的算法对于光线、环境等干扰因素具有较好的鲁棒性，泛化性能好，为基于视频的动作分类识别问题提出了一种可行解决方案。 Motion recognition of table tennis players based on video is studied in this paper. Recognition of human action is challenging in the field of computer vision. Based on videos of ball strike of professional table tennis players against table tennis ball machine, a data set of ball strike of table tennis players is constructed and divided into 5 catalogs of forehand shots, backhand shots, forehand shots, backhand shots and non-stike action. Dense pose of the human body is used to process the constructed data set and extract human body shape from the environment, and then an improved C3D convolutional network is proposed to learn the spatiotemporal features of continuous frames on the data set. Results show that the algorithm proposed in the article has good robustness to interference factors such as light and environment, and good generalization performance, demonstrating a feasible solution to the problem of video-based action classification and recognition.