3D human pose estimation from pseudo-two-viewpoints videos

Human motion analysis using pose estimation has attracted increasing attention in sports science for physical conditioning and skill improvement. This paper presents a method for estimating 3D human poses from RGB videos captured at pseudo two viewpoints. Unlike ordinal approaches, our proposed method requires no specialized equipment such as 3D trackers or multi-view cameras, but only a single consumer RGB camera which enables easy deployment for a wide range of athletes. The subject performs nearly identical motions twice while facing different directions. The two captured videos are treated as pseudo multi-view inputs. The main challenges are the temporal, spatial, and postural inconsistencies that arise between the two recorded videos. We propose a method to address the issues by introducing a head-pose-based camera calibration method, a Dynamic Time Warping (DTW) based temporal alignment, and a pose correction technique grounded in kinematic constraints. After applying these corrections, 3D joint positions are reconstructed via stereo triangulation. Experimental results demonstrate that the proposed method significantly improves 3D pose accuracy compared to state-of-the-art monocular approaches.