关键词: 空间到深度可分离卷积(SPDs-Conv), 加权交并比(WIoU), YOLOv8n, 井下人员检测, 轻量化, 注意力机制, 损失函数

Abstract:

To address the issues of low accuracy and weak robustness in existing detection algorithms due to insufficient lighting, scale differences among personnel, and frequent obstruction by equipment in coal mine environments, as well as the challenges posed by high parameter and computational requirements of some models, which make them difficult to adapt to edge devices underground, an improved YOLOv8n model was proposed to optimize personnel detection tasks in complex mine environments. An enhanced SPDs-Conv module was introduced to enhance the extraction of small target features and improve the recognition accuracy of low-pixel personnel in distant views. Cross stage partial feature fusion + selective kernel attention (C2f_SKAttention) module was designed to strengthen the model's focus on targets of different scales and cope with the scale differences of underground personnel. A dynamic detection head was constructed to adapt to the diversity and complexity of targets, and to improve robustness to occlusion and other scenarios. The WIoU loss function was improved to increase the bounding box localization accuracy and reduce the localization deviation caused by low illumination. The results show that the proposed improved YOLOv8n model achieves an mean average precision (mAP) @0.5 of 83.5% and an mAP@0.5:0.95 of 39.0% on the mine personnel detection dataset. Compared with the original YOLOv8n, the P is improved by 8.5%, the R by 11.9%, the mAP@0.5 by 4.7%, and the mAP@0.5:0.95 by 3.3%. The number of parameters only increases from 3.1M to 3.2M, and the Giga Floating-point operations per second (GFLOPS) rises from 14.0G to 14.4G. The proposed model maintains a lightweight structure while improving detection accuracy and robustness. It effectively alleviates missed detection of small underground targets, insufficient multi-scale adaptation and weak anti-interference capability in complex environments, making it suitable for the limited computing power of underground edge equipment.

Key words: space-to-depth separable convolution (SPDs-Conv), weighted intersection over union (WIoU), YOLOv8n, underground personnel detection, lightweighting, attention mechanism, loss function