3D Human Pose Estimation Based on Temporal Dilated Convolution in the Video

• 1、School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：3D human pose estimation is a hot topic in computer vision. This paper mainly studies the 3D human pose estimation based on video. In order to further improve the recognition accuracy, a 3D human pose estimation algorithm based on temporal convolution is proposed. Firstly, advanced object detection algorithm and 2D human pose estimation algorithm are used to detect the 2D joint position of human body in video frames. Then, the 2D-3D pose lifting network based on temporal convolution proposed in this paper is used to transform 2D human pose into 3D human pose. Compared with the traditional 3D human pose estimation algorithm based on single frame in video, our proposed model can make full use of the temporal information in video to improve the recognition performance, and solve the problem of discontinuity in traditional methods. Experimental results show that the proposed model achieves the best performance on the Human3.6M dataset, and its estimation error is much lower than the existing methods that do not use temporal information and the methods that use RNN to model temporal information for 3D human pose estimation. The experimental results also verify the effectiveness and progressiveness of the proposed model.

Keywords： Computer Vision and Application Human Pose Estimation Temporal Convolution Visualization