基于混合分数引导扩散优化的非标准地形单目人体动作捕捉方法
Monocular Human Motion Capture on Non-Standard Terrain via Hybrid Score-Guided Diffusion Optimization
闫阳旭1
作者信息
- 1. 北京邮电大学计算机学院（国家示范性软件学院），北京100876
- 折叠
摘要
非标准地形（如斜坡、台阶、崎岖路面）下的单目视频人体动作捕捉面临严重的物理失真问题,包括足部穿透地面、脚步滑动及轨迹漂移等。针对现有方法在复杂地形场景下的性能崩塌问题,本文提出了一种基于混合分数引导扩散优化的动作捕捉框架LT-HMR。该框架首先通过运动推断地形提取算法（MITE）从纯运动数据中反演局部地形几何,构建运动-地形交互数据集（MTI-Dataset）；然后设计地形感知运动扩散模型（TAMDiff）学习地形条件下的运动先验；最后在推理阶段引入混合分数引导机制（HS-LTCD）,将扩散先验与地形一致性、时序平滑性等物理约束相结合,实现运动序列的迭代优化。在EMDB、SLOPER4D和RICH三个主流评测集上的实验结果表明,所提方法在根轨迹误差（RTE）、足部滑动（FS）和接触一致性（UCD-Contact）等关键指标上均显著优于现有方法,尤其在非标准地形场景下表现出更强的物理合理性。
Abstract
Monocular human motion capture on non-standard terrain (e.g., slopes, stairs, uneven surfaces) suffers from severe physical artifacts including foot penetration, sliding, and trajectory drift. To address the performance degradation of existing methods in complex terrain scenarios, this paper proposes LT-HMR, a motion capture framework based on hybrid score-guided diffusion optimization. The framework first infers local terrain geometry from pure motion data via the Motion-Inferred Terrain Extraction (MITE) algorithm to construct the Motion-Terrain Interaction Dataset (MTI-Dataset). Then, a Terrain-Aware Motion Diffusion model (TAMDiff) is designed to learn motion priors conditioned on terrain. Finally, a Hybrid Score-guided diffusion optimization mechanism (HS-LTCD) is introduced during inference, combining diffusion priors with physical constraints such as terrain consistency and temporal smoothness for iterative motion optimization. Experimental results on EMDB, SLOPER4D, and RICH datasets demonstrate that the proposed method significantly outperforms existing approaches in key metrics including Root Trajectory Error (RTE), Foot Sliding (FS), and Contact Consistency (UCD-Contact), exhibiting stronger physical plausibility especially on non-standard terrain.关键词
单目动作捕捉/非标准地形/扩散模型/分数引导优化/物理一致性Key words
monocular motion capture/non-standard terrain/diffusion model/score-guided optimization/physical consistency引用本文复制引用
闫阳旭.基于混合分数引导扩散优化的非标准地形单目人体动作捕捉方法[EB/OL].(2026-03-26)[2026-03-28].http://www.paper.edu.cn/releasepaper/content/202603-255.学科分类
计算技术、计算机技术
评论