Scholarlink

<p>Multi-instance pose estimation from depth images or point clouds is a&#xa0;fundamental task and critical challenge in robot navigation and object manipulation, where conventional feature-based methods often suffer from performance degradation under noisy, cluttered, or feature-scarce conditions. To address this limitation, we propose a&#xa0;novel FeatureLess Surface Matching (FLSM) method that eliminates reliance on local or global features/descriptors. FLSM leverages full point relations and pose hashing: first, we derive pose transformation formulas using oriented points to establish model-scene relations; second, we constrain rotational degrees of freedom via discrete angular quantization to generate balanced candidate poses; third, we refine candidates using a&#xa0;4D hashing-based clustering method to reduce computational complexity; finally, we project the 4D pose space onto a&#xa0;2D plane space to enhance the detection capability of pose centers. Experimental validation on the ITODD industrial dataset and real-world multi-instance scenes demonstrates FLSM’s robustness, achieving an Average Recall of 0.589 and outperforming state-of-the-art PPF-based methods by 22.5%. Collectively, FLSM provides a&#xa0;deployable, feature-free solution that enhances the robustness and efficiency of multi-instance pose estimation for industrial and real-world robotics applications.</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

FLSM: Featureless Surface Matching for Multi-instance Pose Estimation

  • 期刊论文

  • 发表日期：

    2025年7月29日

查看全文
  • Huakai Zhao,
  • Jiansheng Li,
  • Xiaomin Liu,
  • Zidi Yang,
  • Xiaojing Dou

摘要

Multi-instance pose estimation from depth images or point clouds is a fundamental task and critical challenge in robot navigation and object manipulation, where conventional feature-based methods often suffer from performance degradation under noisy, cluttered, or feature-scarce conditions. To address this limitation, we propose a novel FeatureLess Surface Matching (FLSM) method that eliminates reliance on local or global features/descriptors. FLSM leverages full point relations and pose hashing: first, we derive pose transformation formulas using oriented points to establish model-scene relations; second, we constrain rotational degrees of freedom via discrete angular quantization to generate balanced candidate poses; third, we refine candidates using a 4D hashing-based clustering method to reduce computational complexity; finally, we project the 4D pose space onto a 2D plane space to enhance the detection capability of pose centers. Experimental validation on the ITODD industrial dataset and real-world multi-instance scenes demonstrates FLSM’s robustness, achieving an Average Recall of 0.589 and outperforming state-of-the-art PPF-based methods by 22.5%. Collectively, FLSM provides a deployable, feature-free solution that enhances the robustness and efficiency of multi-instance pose estimation for industrial and real-world robotics applications.