ScholarWorks@동국대학교

초록

Terrain modeling plays a vital role in applications such as surveying and remote sensing. Point cloud registration, which estimates the rigid transformation to align two frames of point clouds, is a critical step for high-precision spatial modeling in scenarios like smart cities and intelligent transportation. However, conventional methods often suffer from limited perception capabilities and low modeling accuracy in complex environments, making them inadequate for city-scale digital twin applications. To address this challenge, this paper introduces a novel LiDAR point cloud registration framework that leverages multi-scale attention fusion combined with graph optimization. Specifically, the proposed method incorporates a multi-scale attention module that leverages self-attention mechanisms to extract discriminative geometric features and accurately identify key correspondences in low-overlap regions. Subsequently, a pyramid-compatible graph structure is constructed, and a hierarchical maximum clique search algorithm based on distance constraints is employed to eliminate outlier correspondences. Subsequently, a weighted singular value decomposition method is utilized to estimate the transformation matrix from the detected maximum cliques, allowing for precise registration. Experimental results on the KITTI-LC dataset demonstrate that the proposed method improves registration recall by 4.11% and 12.91% over the second-best approach for the challenging distance ranges of 10-20 m and 20-30 m, respectively.

키워드