Rethinking Model Scaling for Ultra-Tiny Person Detection in Open-Pit Mining: Accuracy–Efficiency Trade-off on Edge Hardware

Abstract

Object detection in open-pit mining environments presents a critical yet underexplored challenge: identifying personnel who appear as ultra-tiny targets in surveillance imagery. When workers occupy merely 9×9 pixels on average, far below the 32×32 pixel threshold conventionally used to define small objects, standard detection systems fail to reliably detect them, creating serious safety risks around heavy equipment. This study systematically evaluates nine configurations spanning three state-of-the-art YOLO architectures (YOLO26, YOLOv12, and YOLOv11), each in nano, small, and medium variants, using a proprietary dataset collected at an Indonesian open-pit coal mine. Detection performance was assessed using precision, recall, F1-score, AP@50, and AP@50-95, while deployment feasibility was estimated on the NVIDIA Jetson Nano edge platform. Contrary to the conventional assumption that larger models yield superior accuracy, the results show that nano variants consistently outperform their small and medium counterparts across all three architectures. YOLOv11-Nano achieved the best overall performance, with an F1-score of 0.973, recall of 0.991, and AP@50 of 0.977, while also delivering the fastest edge inference at 31.01 FPS (FP16) and 58.15 FPS (INT8) on the Jetson Nano, comfortably exceeding the 15 FPS real-time threshold. This inverse scaling behaviour is attributed to feature representation collapse in deeper networks, where aggressive downsampling erases the spatial information required to localise ultra-tiny targets. The findings recommend YOLOv11-Nano as a practical, real-time configuration for personnel-detection-based collision avoidance systems in mining vehicles, and highlight architectural scale, rather than capacity, as the key design lever for ultra-tiny object detection at the edge.