自适应采样与几何-空间特征融合的点云配准

打开文本图片集

关键词：点云配准；低重叠率；特征融合；三维重建

中图分类号：TP394.1;TH691.9 文献标识码：A

doi：10.37188/OPE.20253320.3315 CSTR：32169.14.OPE.20253320.3315

Abstract：Point cloud registration in 3D reconstruction scenarios faces significant chalenges，as traditional local feature descriptors often fail due to insufficient keypoints，weak geometric descriptiveness，and poor matching robustness. To address these issues，this study proposed an adaptive sampling and geometryspatial feature fusion algorithm.First，adaptive density-based voxelization folowed by nearest-neighbor downsampling was proposed to address size and density imbalances between low-overlap point cloud pairs while achieving efficient data reduction.Next，surface normals were computed via KD-tree search，and a filtering mechanism incorporating neighborhood point count and linearity constraints was employed to identify salient keypoints. These selected points were subsequently encoded using fused geometry-spatial descriptors to overcome the redundancy and weak descriptiveness of conventional methods. Finally，a bidirectional correspondence approach based on histogram similarity identified reliable point matches，which were then refined through RANSAC to attain robust， high-precision registration under low-overlap conditions.The algorithm was validated on public benchmarks and real-world datasets.Experimental results demonstrate that our method reduces average matching error by 51. 14% ， 64.16% ，and 78% compared to ISS + 3DSC + K4PCS， ISS + FPFH+RANSAC，and TOLDI + RANSAC，respectively. Additionally，our approach achieves the highest runtime efficiency among allcompared methods，evidencing superior accuracy，adaptability，and robustness.

Key words：point cloud registration； low overlap；feature fusion；3D reconstruction

1引言

随着数字化与智能化在工业应用中的快速推进，催生了客观世界三维信息高精度数字化表达的核心需求。（剩余28681字）