Scholarlink

<p>To address the limitations of existing 3D human body reconstruction methods in terms of insufficient precision and coarse detail construction, this paper proposes an optimized solution integrating temporal information, human behavior recognition, and multi-module collaboration. The algorithm centers on pixel-aligned hidden functions to establish a full-process reconstruction framework encompassing "temporal constraints-behavioral guidance-feature selection-attitude optimization-3D mapping-detail enhancement". By employing the SAD algorithm for optimal matching point selection and LSTM for temporal dependency capture, the algorithm achieves cross-frame feature coordination. Subsequently, the CNN-LSTM model performs human behavior recognition, using behavioral categories to guide SMPL model’s attitude parameter prior and attitude discriminator constraints. Posture normalization eliminates individual variations, while the integration of SMPL model and PIFuHD hidden function enables structured 3D mapping. Finally, octree acceleration grids are utilized to output high-precision 3D human models. Experimental results demonstrate that the proposed algorithm outperforms traditional and literature methods, achieving stable human detail construction in both static standing scenarios and dynamic running scenes.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-module collaborative 3D human body modeling algorithm based on PIFuHD

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月2日

查看全文
  • Zemin Qiu,
  • Jiajun Zou,
  • Shaojiang Liu,
  • Feng Wang

摘要

To address the limitations of existing 3D human body reconstruction methods in terms of insufficient precision and coarse detail construction, this paper proposes an optimized solution integrating temporal information, human behavior recognition, and multi-module collaboration. The algorithm centers on pixel-aligned hidden functions to establish a full-process reconstruction framework encompassing "temporal constraints-behavioral guidance-feature selection-attitude optimization-3D mapping-detail enhancement". By employing the SAD algorithm for optimal matching point selection and LSTM for temporal dependency capture, the algorithm achieves cross-frame feature coordination. Subsequently, the CNN-LSTM model performs human behavior recognition, using behavioral categories to guide SMPL model’s attitude parameter prior and attitude discriminator constraints. Posture normalization eliminates individual variations, while the integration of SMPL model and PIFuHD hidden function enables structured 3D mapping. Finally, octree acceleration grids are utilized to output high-precision 3D human models. Experimental results demonstrate that the proposed algorithm outperforms traditional and literature methods, achieving stable human detail construction in both static standing scenarios and dynamic running scenes.