Multi-Branch High-Dimensional Guided Transformer-Based 3D Human Posture Estimation

Abstract

The human pose paradigm is estimated using a transformer-based multi-branch multidimensional directed the three-dimensional (3D) method that takes into account self-occlusion, badly posedness, and a lack of depth data in the per-frame 3D posture estimation from two-dimensional (2D) mapping to 3D mapping. Firstly, by examining the relationship between the movements of different bones in the human body, four virtual skeletons are proposed to enhance the cyclic constraints of limb joints. Then, multiple parameters describing the skeleton are fused and projected into a high-dimensional space. Utilizing a multi-branch network, motion features between bones and overall motion features are extracted to mitigate the drift error in the estimation results. Furthermore, the estimated relative depth is projected into 3D space, and the error is calculated against real 3D data, forming a loss function along with the relative depth error. This article adopts the average joint pixel error as the primary performance metric. Compared to the benchmark approach, the estimation findings indicate an increase in average precision of 1.8 mm within the Human3.6M sample.