基于倒置残差的井下无人车目标检测研究  

Research on Target Detection of Underground Unmanned Vehicle Based on Inverted Residual

作　　者：杨伟康 吕文生[1] 杨鹏[2] 张宇栋 党龙飞 YANG Weikang;LYU Wensheng;YANG Peng;ZHANG Yudong;DANG Longfei(School of Civil and Resources engineering,University of Science and Technology Beijing,Beijing 100083,China;Collage of Urban Rail Transit and Logistics,Beijing Union University,Beijing 100101,China)

机构地区：[1]北京科技大学土木与资源工程学院,北京100083 [2]北京联合大学城市轨道交通与物流学院,北京100101

出　　处：《矿业研究与开发》2024年第4期222-227,共6页Mining Research and Development

基　　金：国家重点研发计划项目(2021YFC3001300)。

摘　　要：对井下光线昏暗、光照不均、背景复杂等特殊工况环境的影响下,所形成的图像目标存在细节特征少、图像模糊等问题进行了研究,提出了一种基于倒置残差结构改进YOLOv5s模型的检测算法,以此解决井下目标检测精度低的问题。首先主干网络部分引入通道注意力神经网络模块(SE-Net),提高检测精度;颈部网络部分在BottleneckCSP模块中引入倒置残差结构,将通道进行扩充,丰富特征数量,进一步提升检测精度。在自建井下数据集上进行检测试验,结果表明,基于倒置残差的模型平均检测精度均值(交并比为0.5)达84.4%,相比YOLOv5s模型精度提高了16.7个百分点,参数量减少了17.1%,模型轻量化且精度高,可有效改善井下目标检测精度低的问题,基本满足井下无人车目标检测的需求。Under special working conditions such as dim underground light,uneven lighting,and complex background,the image targets have problems of few detailed features and blurred images.Thus,a detection algorithm based on the inverted residual structure improved YOLOv5s model was proposed to solve the problem of low accuracy in underground target detection.Firstly,the backbone network introduced a Squeeze and Excitation Networks(SE Net)module to improve detection accuracy.The neck network introduced an inverted residual structure in the BottleneckCSP module to expand the channels and enrich the number of features,further improving detection accuracy.The detection experiment was conducted on a self-built underground dataset,and the results show that the average detection accuracy of the model in this paper(intersection over union is 0.5)reached 84.4%,which improved the accuracy of the YOLOv5s model by 16.7 percentage points and reduced the number of parameters by 17.1%.The model in this paper is lightweight and has a high accuracy,which can effectively improve the problem of low accuracy in underground target detection and basically meet the needs of underground unmanned vehicle target detection.

关 键 词：井下图像 目标检测 倒置残差 注意力 YOLOv5s模型 

分 类 号：TD529[矿业工程—矿山机电] TP29[自动化与计算机技术—检测技术与自动化装置]