基于SCI-YOLOv8的井下无人电机车低照度目标检测技术  

Low Light Target Detection Technology for Underground Unmanned Electric Vehicles Based on SCI-YOLOv8

作　　者：雷杨 何江 秦丽杰 何文轩 纪雅溟 柳小波 LEI Yang;HE Jiang;QIN Lijie;HE Wenxuan;JI Yaming;LIU Xiaobo(China Mineral Resources Group Co.,Ltd.,Beijing 100142,China;School of Civil and Resource Engineering,University of Science and Technology Beijing,Beijing 100083,China;Liaoning Metallurgical Geology 402 Team Co.,Ltd.,Anshan 114001,China;Ansteel Group Mining Co.,Ltd.,Anshan 114001,China;Institute of Smart Mining,University of Science and Technology Liaoning,Anshan 114001,China)

机构地区：[1]中国矿产资源集团有限公司,北京100142 [2]北京科技大学土木与资源工程学院,北京100083 [3]辽宁省冶金地质四〇二队有限责任公司,辽宁鞍山114001 [4]鞍钢集团矿业有限公司,辽宁鞍山114001 [5]辽宁科技大学智慧矿山研究院,辽宁鞍山114001

出　　处：《金属矿山》2025年第2期172-179,共8页Metal Mine

基　　金：国家自然科学基金项目(编号:52474172)。

摘　　要：在井下电机车无人驾驶障碍物检测任务中,由于光照不均、粉尘等因素的影响,电机车障碍物图像目标检测系统会出现提取目标特征困难、目标识别精度降低等问题,易导致障碍物误检和漏检。针对上述问题,提出一种基于SCI-YOLOv8的低照度目标检测算法,将SCINet自校正照明网络与YOLOv8目标检测算法相融合,使YOLOv8算法更有利于低光照目标检测。并将LSKA注意力机制嵌入到YOLOv8网络中Head部分C2f的末端,降低计算和内存成本的同时,保持了高效的图像处理能力。实验结果表明,本研究所提出的算法在公共低光数据集Exdark目标检测中,mAP@50为57.7%,mAP@50∶95为35.4%。相较于原始YOLOv8目标检测算法mAP@50提高了1个百分点,mAP@50∶95提高了1.4个百分点。在井下低光数据集LLP目标检测中,SCI-YOLOv8模型的mAP@50达到97.3%,mAP@50∶95为68.2%,相较于原始模型分别提高了3.4个百分点和8.6个百分点。本研究所提出的SCI-YOLOv8算法在低光场景的目标检测中具有优越性,能满足井下低光场景的目标检测任务要求,为井下电机车安全、高效、智能运行提供了技术支撑。In the obstacle detection task of unmanned driving of underground electric locomotives,due to the influence of uneven illumination,dust and other factors,the obstacle image target detection system of electric locomotive will have problems such as difficulty in extracting target features and reducing the accuracy of target recognition,which can easily lead to false detection and missed detection of obstacles.In order to solve the above problems,a low-illumination target detection algorithm based on SCI-YOLOv8 was proposed,which fused the SCINet self-correcting illumination network with the YOLOv8 target detection algorithm,so that the YOLOv8 algorithm was m ore conducive to low-illumination target detection.The LSKA attention mechanism is embedded at the end of the C2f of the Head part of the YOLOv8 network,which reduces the computational and memory costs while maintaining efficient image processing capabilities.Experimental results show that the proposed algorithm has a mAP@50 of 57.7% and a mAP@50∶95 of 35.4% in the public low-light dataset Exdark.Compared with the original YOLOv8 object detection algorithm mAP@50 it is 1 pecent point higher,and mAP@50∶95 is improved by 1.4 pecent point.In the LLP object detection of the downhole low-light dataset,the mAP@50 of the SCI-YOLOv8 model reached 97.3% and the mAP@50:95 was 68.2%,which were 3.4 pecent point and 8.6 pecent point higher than the original model,respectively.The SCI-YOLOv8 algorithm proposed in this paper has superiority in target detection in low-light scenes,which can meet the requirements of target detection tasks in low-light underground scenes,and provides technical support for the safe,efficient and intelligent operation of underground locomotives.

关 键 词：YOLOv8 行人检测 注意力机制 低照度 

分 类 号：TD676[矿业工程—矿山机电]