三维点云表示的人体动作序列预测

Human Action Sequence Prediction of 3D Point Cloud Representation

目前对三维人体动作序列的预测工作相对较少,且主要使用三角形网格表示人体模型,不如三维点云那样简单又容易获取。为此,该文用三维点云表示人体模型,提出一种基于MeteorNet的点云动作序列预测方法。将动作序列中不同时刻的三维点云融合在一起,寻找点的时空邻域进行分组;叠加三层Meteor模块在时空邻域聚合信息,以获取点云序列的时空特征;通过三层全连接网络预测动作的点云坐标。实验结果表明,该方法预测出的人体动作与真实动作的误差较小。

Few works on action prediction of 3D human have been reported,and most of them represent human model with triangular mesh,which is not as simple and obtainable as 3D point clouds.Therefore,this paper proposes a point cloud action sequence prediction method based on MeteorNet by using 3D point clouds to represent human model.In an action sequence,the 3D point clouds at different times are fused together for finding spatiotemporal neighborhoods of the point clouds and grouping them;Three-layer Meteor modules are superimposed in the spatiotemporal neighborhoods for aggregating information and obtaining spatiotemporal features of the point cloud sequence;thus,the point cloud coordinates of action are predicted by a three-layer fully connected network.Experimental results show that the human actions predicted by the proposed method have lower errors with real actions.