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

Abstrak

Object detection in open-pit mining environments presents a critical yet underexplored challenge: identifying personnel who appear as ultra-tiny targets in imagery captured from mining vehicles and surveillance systems. When persons occupy merely 9×9 pixels on average far below the 32×32 pixel threshold conventionally defining small objects standard detection architectures exhibit severe performance degradation. This study presents a systematic comparative evaluation of nine model configurations spanning three contemporary YOLO architectures (YOLO26, YOLOv12, and YOLOv11) across nano, small, and medium variants for ultra-tiny person detection in real mining imagery. Our experiments reveal a counterintuitive finding: nano variants consistently outperform their larger counterparts across all architectures, challenging prevailing assumptions about model scaling in deep learning. YOLOv11-Nano achieved the highest performance with 0.973 F1-score and 0.991 recall while maintaining 31 frames per second on NVIDIA Jetson Nano hardware comfortably exceeding real-time requirements for collision avoidance systems. The attention mechanisms in YOLOv12, despite theoretical promise for preserving spatial information, introduced computational overhead without accuracy benefits at ultra-tiny scales. These findings suggest that aggressive downsampling in deeper networks eliminates spatial information critical for ultra-tiny detection before meaningful feature encoding occurs. The results provide practical guidance 1 for deploying person detection systems in mining safety applications and contribute empirical evidence that conventional model scaling assumptions may not hold at extreme object scales.

Identifier

• DOI: https://doi.org/10.36227/techrxiv.177160646.64552622/v1
• OpenAlex: https://openalex.org/W7130692088