Adaptive Motion-aware Body-surface Evidence Representation for Lidar-based Human Motion Capture

LiDAR-based human motion capture aims to recover temporally coherent 3D human poses from sparse, occluded, and noisy point-cloud observations. Although complementary cross-frame observations can alleviate single-frame incompleteness, existing approaches typically organize temporal evidence through estimated skeletons, making cross-frame association vulnerable to skeletal estimation uncertainty. This limitation becomes particularly severe under sparse measurements and self-occlusion, as unreliable skeletal estimates propagate errors during temporal reasoning. In this paper, we propose a new framework called Adaptive Motion-aware Body-surface Evidence Representation (AMBER) to decouple temporal evidence organization from skeletal reasoning. Specifically, AMBER first constructs a faithful representation by organizing multi-scale cross-frame evidence around adaptive surface anchors. The resulting representation is further incorporated into a Geometry-Motion Refinement (GMR) module to refine the corresponding joint positions and motion dynamics in the skeletal domain. In this way, rather than prematurely compressing evidence into sparse joints, AMBER retains rich geometric details and motion cues on the body surface to ensure robust cross-frame association. Extensive experiments on multiple LiDAR benchmarks demonstrate that AMBER consistently outperforms SOTA methods.