Fusion-Based Semantic Segmentation and 3D Reconstruction Using Radar–LiDAR Point Clouds: A Comparative Evaluation of DeepLabv3 and FCN-ResNet Against Traditional Architectures

Reliable person segmentation with sparse 3D sensors degrades significantly under adverse atmospheric conditions. This work presents a controlled comparative evaluation of four segmentation architectures—U-Net, Mask R-CNN, DeepLabV3+, and FCN-ResNet—on a fused Radar–LiDAR dataset for binary person–background segmentation and applies a dual-domain evaluation procedure that formally links 2D pixel-level overlap (IoU, Dice) to 3D geometric fidelity (Chamfer distance, Completeness) through mask back-projection onto fused point clouds. Raw point clouds are rasterized into range–intensity grids enriched with Radar reflectivity; the predicted masks are then reprojected into 3D space and evaluated using Chamfer distance and Completeness under three controlled visibility conditions. U-Net achieves the highest 2D overlap (IoU = 0.82, Dice = 0.89), while DeepLabV3+ delivers the best 3D reconstruction fidelity (Chamfer = 0.021 m, Completeness = 93.4%) and the highest overall accuracy (97.9%). This dissociation between 2D overlap and 3D fidelity is explained by DeepLabV3+’s multi-scale Atrous Spatial Pyramid Pooling (ASPP), which reduces boundary fragmentation during back-projection; more than 70% of the Chamfer deviation across competing architectures originates at object contours. Mask R-CNN performs well when instances are clearly separated, and FCN-ResNet offers the lowest computational cost at reduced boundary precision. Radar–LiDAR fusion sustains an IoU within 3% of clear-weather performance under dense fog, whereas LiDAR-only inputs degrade by more than 12%. Due to the 12:1 background-to-person class imbalance, overlap-based metrics (IoU, Dice) are prioritized over raw accuracy in all reported comparisons. These results provide actionable deployment guidance and constitute a reproducible evaluation procedure for future sparse-sensor fusion studies, independently of the architectures evaluated.