Point Cloud Change Detection Based on ROI Similarity

Abstract

Laser point cloud change detection is susceptible to spatial–structural uncertainties induced by residual registration errors, uneven point cloud density, viewpoint shift and object occlusion, which frequently produce numerous false alarms of spurious changes. To address this issue, this paper proposes an ROI-level 3D point cloud change detection method based on keypoint similarity. Potential change points are first filtered according to point-to-point distances, and Regions of Interest (ROIs) are generated via RBNN clustering. Global point cloud comparison is converted into local comparison among ROI units. Keypoints within ROIs are extracted for surface reconstruction, and effective overlapping regions between paired ROIs are obtained. Regional similarity is constructed by fusing surface normal consistency and dynamic-evolution features of complex networks for change discrimination. Experiments on simulated mine-tunnel data and field-measured underground-garage data demonstrate that, under unseen background disturbances including residual registration error, occlusion and uneven point cloud density, the proposed method achieves the highest F1-score of 0.9004 for the simulated tunnel scenario compared with competing methods, and delivers effective false-alarm suppression without degrading detection recall for real-change regions.