Abstract
Terrestrial Laser Scanning (TLS) is pivotal for creating high-precision 3D geospatial data in urban digital twins and Scan-to-BIM applications. However, TLS point clouds frequently suffer from severe glass-induced reflection artifacts, which fundamentally compromise geometric fidelity and downstream automated modeling. Existing methods struggle with real-world physical complexities because they rely heavily on idealized reflection symmetry assumptions and fail to address the critical “measurement void” problem caused by transparent surfaces. To address these limitations, we propose a unified cross-modal framework RADAR for robust reflection artifact removal in complex urban environments. First, our approach leverages a multimodal vision foundation model—integrating panoramic imagery and LiDAR-derived spatial maps—to accurately detect glass regions. These initial masks are then geometrically refined and completed via ray-casting to recover missing surfaces. Second, to handle imperfect reflections, we introduce the Reflection-aware Local–Global Geometric Similarity (RE-LGGS) descriptor. By jointly encoding multi-scale PCA-derived structures and spatial orientation consistency, RE-LGGS robustly identifies virtual points even under severe geometric distortion and data sparsity. Extensive evaluations on multiple public TLS datasets demonstrate that the proposed framework achieves competitive performance across diverse glass-scene conditions.