Abstract: In real-world scenarios, human images are often affected by clothing, self-occlusion, and environmental object occlusion, leading to insufficient visible information. Existing human mesh recovery methods tend to degrade to average models in shape estimation, failing to faithfully recover individual geometric characteristics. To address this issue, Human Dimension Attention Regressor for monocular occluded human mesh recovery (HDAR) was proposed for occluded human mesh recovery. Firstly, the visible human regions were used to infer the dimensions of occluded parts. Secondly, a hierarchical proportion constraint mechanism was then introduced, in which first-level constraints were applied to adjacent human parts and second-level constraints were applied to more distant parts, ensuring that the regressed shapes conform to human anatomical structure. Finally, 2D joint information was integrated with the estimated human dimensions in an iterative optimization process to improve pose estimation accuracy. On the 3DPW (3D Poses in the Wild) dataset, the method achieves a per-vertex error (PVE) of 65.2 mm, which is 10.7 mm lower than that of Multi-HMR(Multi-Person Whole-Body Human Mesh Recovery) under occlusion conditions, corresponding to a 14.1% error reduction. Visualization experiments demonstrate that the method effectively improves the recovery accuracy of human shape and pose in complex occlusion scenarios.

Key words: Keywords: monocular human recovery, human dimension, human measurement, occluded human body, hierarchical proportion constraint, iterative optimization

摘要： 在真实场景中人体图像常受服装、自身姿态及环境物体的遮挡，导致可见信息不足，使现有人体重建方法在形状 建模上易退化为均值模型，难以真实还原个体真实特征。针对这一问题，提出一种基于人体尺寸信息的遮挡人体网格重建方 法用于单目遮挡人体网格恢复的人体尺寸注意力回归方法(Human Dimension Attention Regressor, HDAR)。该方法首先利用可见 区域的人体尺寸推理被遮挡部分的尺寸信息；其次引入人体维度的分级比例约束，在临近部位间建立一级约束、在较远部位 间建立二级约束，使回归形状符合人体结构特征；最后结合二维关节点信息与人体尺寸进行迭代优化，提升姿态估计精度。 实验结果表明，在 3DPW(3D Poses in the Wild)数据集上，该方法的逐顶点误差 PVE(Per Vertex Error)为 65.2 mm，相比 Multi-HMR(Multi-Person Whole-Body Human Mesh Recovery)减少了 10.7 mm，减少了 14.1%的误差。可视化实验表明，所提方 法能够在复杂遮挡场景下有效提升人体形状与姿态的重建精度。

关键词: 单视图人体重建, 人体尺寸, 人体测量, 遮挡人体, 分级比例约束, 迭代优化