文献类型：期刊文献

中文题名：面向矿下无人车的红外小目标人员检测算法

英文题名：Infrared small target pedestrian detection algorithm for underground unmanned vehicles

作者：范杨杨[1,2];刘元盛[1,2];王庆闪[3]

第一作者：范杨杨

机构：[1]北京联合大学北京市信息服务工程重点实验室,北京100101;[2]北京联合大学机器人学院,北京100101;[3]中汽研(天津)汽车工程研究院有限公司,天津300300

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2025

卷号：44

期号：8

起止页码：133-137

中文期刊名：传感器与微系统

外文期刊名：Transducer and Microsystem Technologies

收录：;北大核心:【北大核心2023】；

基金：国家自然科学基金资助项目(62371013);北京市属高等学校高水平科研创新团队建设支持计划项目(BPHR20220121)。

语种：中文

中文关键词：矿下环境;无人巡检车;YOLOv11n;可编程梯度信息;红外小目标检测

外文关键词：underground mining environment;unmanned inspection vehicle;YOLOv11n;PGI;infrared small target detection

摘要：矿下环境主要依赖强光源照明,导致无人巡检车使用可见光传感器采集的图像因过度曝光而丢失人员目标。红外传感器虽能适应此类光照条件,但其成像分辨率较低,加之小目标特征不显著、目标密集等因素,导致人员漏检率较高。针对上述问题,本文提出基于YOLOv11n的红外小目标人员检测算法YOLO-PDL。首先,采用可编程梯度信息(PGI)策略重构主干网络,增强网络对矿下人员的表征能力,减少精度损失;其次,在C3K2模块中嵌入扩张残差(DWR)模块,增强网络骨干的特征提取能力;最后,在颈部网络中添加大型可分离核注意力(LSKA)机制,提升网络对多尺度目标的适应性和准确度。实验结果表明:所提算法在自建矿下红外数据集和FLIR数据集上的mAP分别提升3.8%和2.7%,显著提升了模型检测性能。
Underground mining environments primarily rely on strong artificial lighting,causing images captured by visible light sensors on unmanned inspection vehicles to suffer from overexposure,resulting in the loss of personnel targets.Although infrared sensors can adapt to such lighting conditions,their lower imaging resolution,combined with factors such as insignificant small target features and dense target distribution,leads to a high miss rate for personnel detection.To address these issues,this paper proposes YOLO-PDL,an infrared small target personnel detection algorithm based on YOLOv11n.Firstly,the programmable gradient information(PGI)strategy is used to reconstruct the backbone network,enhancing the network’s ability to represent underground personnel and reducing accuracy loss.Secondly,the dilation wise residual(DWR)module is embedded into the C3K2 module to enhance the feature extraction capability of the network backbone.Finally,the large separable kernel attention(LSKA)mechanism is added to the neck network,improving the network’s adaptability and accuracy for multi-scale targets.Experimental results show that the improved algorithm achieves mAP improvements of 3.8%and 2.7%on the self-built underground mining infrared dataset and the FLIR dataset,respectively,significantly enhancing the model’s detection performance.