Robust distance perception for construction machinery via spatial-semantic LiDAR-camera fusion

Reliable distance perception for intelligent construction equipment in unstructured work sites is challenged by illumination variation, occlusion, and background clutter, which hinder LiDAR–camera feature association. We propose a spatial-semantic association framework integrating visual detection, point-cloud processing, and cross-modal matching. The visual branch employs a lightweight scalable detection head (SS-Head), with EIoU and Focal Loss to improve localisation in dense scenes and hard-sample recall. The point-cloud branch combines multistage-optimised generalised iterative closest point (GICP) registration with PointNet++ semantic segmentation in a cascaded architecture to refine alignment and extract target-level three-dimensional semantics. A dual-layer R-Tree replaces exhaustive frustum-based matching with hierarchical spatial retrieval for associating detection boxes and semantic point-cloud clusters. On an RK3588 embedded platform, the improved visual model achieves 65.69% mAP@0.5:0.95, while registration RMSE under low-overlap conditions falls to 0.013 m. Across evaluated low-illumination, 30%–50% occlusion, and multi-target scenarios at 5–25 m, the valid ranging rate is at least 85%, with an overall MAE of 0.06 m, RMSE of 0.10 m, and average processing time of 147.2 ms per frame. These results indicate potential for safety warnings on resource-constrained construction equipment.