Voxel-based recovery and trajectory separable error rectification for multi-view 3D pose reconstruction of jump analysis in figure skating

Abstract

Successful jumps in figure skating with critical parameters such as proper jump height, spin speed, and the number of jump rotations, which are valuable for analysis in athlete training. Driven by recent computer vision applications, reconstructing 3D poses of the athlete in figure skating to extract the significant variables has become increasingly important. However, a large number of conventional works have obtained 3D poses from corresponding 2D information directly, which ignores the uniqueness of figure skating, such as self-occlusion, abnormal poses, etc. This paper proposes a multi-view voxel based system for calibration and error correction to reconstruct the 3D jumping poses of figure skaters in the presence of 2D heatmaps. The proposed method consists of two key components: Voxels based recovery method of high probability area in 2D heatmap; Plain 2D smoothness and motion trajectory and relative joint positions separable 3D smoothness based rectification method. This work is proven to be applicable to 3D pose dynamics in figure skating jumping motion. Mean Per Joint Position Error (MPJPE) is: 34.58mm in the pre-jump stage, 16.51mm in the jumping stage, 15.73mm in the post-jump stage, and 16.93mm in the whole jump stage, which is 36% improvement compared with the conventional work.