Abstract
In complex urban environments, traversable spaces beneath bridges and roadside trees are often obscured in UAV aerial imagery. To address this challenge, we propose a novel end-to-end Occlusion-Informed 3D Terrain Mapping (OI-3DTM) framework that enables accurate voxel-level modeling of hidden structures, providing a refined spatial foundation for traversability analysis. Existing methods largely rely on 2D semantic completion or expensive multi-view and LiDAR-based mapping, which either lack sufficient 3D expressiveness or fail to meet real-time and resource constraints. OI-3DTM integrates self-supervised monocular depth estimation with semantic segmentation and 3D reconstruction to explicitly infer traversable voxel structures in occluded scenes. A unified encoder–decoder network fuses RGB imagery with prior depth features, enhanced by a semantic-partitioned hybrid loss and confidence-ranked depth consistency, thereby improving depth accuracy in textureless or geometrically discontinuous regions. Furthermore, we introduce a geometry prior-informed voxel reasoning module that employs parameterized cavity curves to model the underlying spatial geometry of occluded regions, enabling the reconstruction of invisible structures. We evaluate OI-3DTM on three benchmark datasets for depth estimation and conduct 3D inference experiments on the UAV-LADIS dataset. Experimental results demonstrate that the proposed method consistently outperforms existing approaches in both depth estimation and 3D spatial reasoning, achieving a 61% improvement in IoU and reducing the cross-entropy error to 0.116 in occlusion scenarios. These results indicate that OI-3DTM significantly enhances occlusion-aware spatial perception and exhibits strong potential for practical deployment in UAV-based urban navigation applications.
Funding Information
  • National Natural Science Foundation of China (62173220)
  • National Natural Science Foundation of China (62303294)
  • Jiangxi Research Institute, Beihang University