Intelligent Visual Detection of Surrounding Personnel for Underground Load–Haul–Dump Vehicles in Complex Mining Environments

In underground mining environments, collisions between Load–Haul–Dump (LHD) machines and personnel pose serious safety risks. To address challenges such as uneven illumination, dust interference, occlusion, and limited computational resources, this paper proposes FP-DETR, an improved RT-DETR framework for underground personnel detection. The model integrates a CSP-FFCM module for efficient spatial–frequency feature extraction, a polarity-aware linear attention-based POTE module for enhanced global feature interaction, and a Matching-Aware Loss to improve confidence–quality alignment and reduce false alarms. Experiments on a self-constructed dataset show that FP-DETR achieves 88.5% mAP@0.5 and 58.6% mAP@0.5:0.95, improving RT-DETR-r18 by 1.5% and 3.5%, respectively, while reducing parameters from 20.0 M to 15.8 M and GFLOPs from 57.3 to 51.2. Field tests show a reduction in frame-level FPR from 4.1% to 3.4%. However, validation is limited to no-person scenarios, and full operational evaluation is required for comprehensive assessment.