Mask-Guided Frequency-Aware Sparse Elevation Map Completion for Off-Road Vehicle Terrain Perception

Sparse sampling and occlusion limit geometric analysis of LiDAR-derived elevation maps. This study proposes MGFE-Net, a mask-guided and frequency-aware network for elevation-map completion. The method combines elevation, an observation mask, selected-return count, and height variance to represent surface geometry, measurement support, and height dispersion. A pooling-based enhancement module integrates coarse contextual features with learned residual details within a U-Net-like encoder–decoder, while mask-based backfilling retains observed input elevations. Experiments use spatially separated training, validation, and test regions of RELLIS-3D sequence 00000, with exclusion gaps and disjoint source frames. Four network variants are trained from scratch using three random seeds and evaluated against temporally screened multi-frame reference surfaces. MGFE-Net achieves a root mean square error of 0.1593 ± 0.0056 m and a gradient error of 0.1427 ± 0.0024 m/m, reducing the Base U-Net means by 6.77% and 7.95%, respectively. The completed maps increase the computability of slope, step height, and surface roughness, and result in lower mean reference-based feature errors among the evaluated variants. However, exploratory risk evaluation still identifies false-low classifications and spurious connections. These findings support improved reconstruction within the tested spatial split and reference protocol, while broader generalization and safe vehicle traversability remain unverified.