Non-rigid registration of continuous-frame human point clouds fusing body priors

Abstract

Continuous-frame human point cloud registration is a fundamental task for dynamic three-dimensional human reconstruction, motion analysis, and digital-human modeling. Affected by articulated motion and locally similar structures, conventional rigid and non-rigid registration methods often suffer from mismatches, collapse, and unnatural stretching on this task. Building on the occlusion-aware non-rigid registration method OAReg as the backbone, this paper proposes an unsupervised non-rigid registration method for continuous-frame human point clouds, termed OAReg-Human. The method takes two adjacent frames as the source and target clouds, uses a sine-activation implicit network (SIREN) to represent a continuous displacement field, and achieves occlusion-robust alignment through a correntropy Chamfer distance and a locally linear reconstruction constraint. Body-part labels are introduced as a structural prior, with a part-smoothness loss and a part-rigidity loss constraining intra-part deformation, and a per-part Kabsch rigid refinement is added as post-processing to stabilize local regions such as the limbs and torso. Experiments on the public CAPE dataset show that the method outperforms ICP, CPD, and OAReg in end-point error, matching accuracy, and outlier ratio, verifying the effectiveness of the body-part prior and per-part rigid refinement.