A Density Clustering and Dynamic Object Filtering Framework for Robust LiDAR Registration

Le résumé fourni par la source

Achieving robust point cloud registration in dynamic environments remains challenging. Conventional voxel downsampling often leads to loss of geometric information and shows limited adaptability to complex scenes. To address this, this paper proposes a robust registration framework that combines density-based clustering and dynamic object removal. Ground segmentation is first performed to eliminate ground influence; non-ground points are then clustered using DBSCAN, and dynamic objects are identified and removed via frame-graph cluster association to obtain a relatively static environmental point cloud. On this basis, the point cloud is divided into two categories: high-density and sparse. Point-to-plane ICP is applied to highdensity points to fully exploit local planar constraints and improve convergence stability; point-to-point ICP is applied to sparse points to preserve edges and fine structures, while down-weighting and outlier-suppression strategies are used to reduce noise interference. The two registration results are finally fused in a unified pose estimation, achieving more stable convergence and higher localization accuracy across diverse scenarios. Experimental results show that the proposed method attains higher accuracy and robustness than voxel-grid baselines on public datasets and real-vehicle tests, with particularly strong performance in dynamic urban scenes.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.