LL-YOLO: a personnel detection algorithm for low-light environment in coal mines

A personnel detection algorithm LL-YOLO （Low-Light YOLO） is proposed to address the problems of blurred boundaries and degraded details of personnel targets caused by insufficient lighting and dust scattering in low-light environments underground coal mines. Firstly, a CondConv-based dynamic convolution is employed to enhance the backbone network, improving sensitivity to contours and texture details in blurred boundary regions. Secondly, a recalibrated feature pyramid network is designed to optimize the original feature fusion structure, thereby strengthening spatial alignment and representational consistency across multi-scale features. Thirdly, a lightweight shared detection head is introduced, which reduces model parameters while enhancing the perception of local structures in weakly visible regions. Finally, a localization regression loss based on normalized Wasserstein distance is incorporated to improve stability and accuracy in detecting targets with blurred boundaries and structural degradation under low-light conditions. The experimental results on the self built coal mine low-light personnel detection LLMiners dataset, as well as the low-light datasets LLVIP and ExDark, demonstrate that LL-YOLO achieves improvements of 2.28%, 2.09%, and 3.25% in mAP@0.50:0.95 compared with the baseline YOLOv11s. Moreover, when compared with state-of-the-art models such as Faster R-CNN, YOLO series, and RT-DETR series, LL-YOLO consistently outperforms them across multiple metrics. In addition, compared with mainstream models such as Faster RCNN, YOLO series, and RT-DETR series, LL-YOLO also has advantages in multiple indicators. LL-YOLO effectively improves the personnel detection performance in low-light scenarios of coal mines while controlling computational costs, providing reliable technical support for intelligent safety systems such as real-time monitoring of underground personnel and abnormal behavior warning.