测绘通报 ›› 2026, Vol. 0 ›› Issue (10): 127-132.doi: 10.13474/j.cnki.11-2246.2026.1019

• 技术交流 • 上一篇    

改进YOLO11的铁路无人机巡检作业人员防护装备检测方法

宋宗莹1, 曾杉2, 张正军3, 尹太军3   

  1. 1. 中国神华能源股份有限公司, 北京 100011;
    2. 中国科学院地理科学与资源研究所, 北京 100101;
    3. 中科吉芯(秦皇岛)信息技术有限公司, 河北 秦皇岛 066000
  • 收稿日期:2026-06-18 发布日期:2026-11-06
  • 通讯作者: 曾杉。E-mail:zs@lreis.ac.cn
  • 作者简介:宋宗莹(1981—),男,博士,教授级高级工程师,研究方向为智慧铁路、通信信号。E-mail:10000372@ceic.com
  • 基金资助:
    国家能源集团科技创新项目(GJNY-24-153);重载铁路数字孪生装车站智能管控技术研究与应用示范(KF2025-044)

Personal protective equipment detection for railway workers in UAV inspection using an improved YOLO11

Song Zongying1, Zeng Shan2, Zhang Zhengjun3, Yin Taijun3   

  1. 1. China Shenhua Energy Company Limited, Beijing 100011, China;
    2. Institute of Geographic Sciences and Natural Research, Chinese Academy of Sciences, Beijing 100101, China;
    3. Digitwinology International (Qinhuangdao)Co., Ltd., Qinhuangdao 066000, China
  • Received:2026-06-18 Published:2026-11-06

摘要: [目的] 针对铁路无人机影像中安全帽、反光衣等防护装备目标尺度小、特征信息易丢失且受复杂背景干扰而导致漏检的问题,本文提出了一种融合人体空间结构先验与浅层特征增强的人员防护装备检测方法。[方法] 该方法基于YOLO11n模型,依据安全帽、反光衣与人员目标之间的空间依附关系构建先验区域,并设计区域加权分类损失,引导模型增强人体合理区域内防护装备目标的学习;同时引入高分辨率浅层特征分支,提升小尺度目标表征能力。[结果] 自建数据集上,本文方法的P、R、mAP 0.5、mAP 0.5:0.95分别达92.3%、89.2%、94.3%、50.7%,较基线分别提升8.6、9.1、8.7和9.0个百分点。[结论] 本文方法有效提高了铁路无人机影像中人员及防护装备的检测精度,为施工安全定位与线路安全管理提供了支撑。

关键词: 无人机影像, 无人机巡检, 人员防护装备检测, 人体空间结构先验, 小目标检测

Abstract: [Purposes] To address missed detections caused by the small size,weak feature representation,and complex-background interference of safety helmets,reflective vests,and other personal protective equipment in railway UAV imagery,this paper proposes a detection method integrating human spatial-structure priors and shallow feature enhancement. [Methods] Based on YOLO11n,prior regions are constructed according to the spatial relationships between personnel,safety helmets,and reflective vests.A region-weighted classification loss is designed to strengthen the learning of protective equipment within anatomically reasonable regions.Meanwhile,a high-resolution shallow feature branch is introduced to improve small-object representation. [Findings] On the self-built dataset,the proposed method achieved a P of 92.3%,R of 89.2%,mAP 0.5 of 94.3%,and mAP 0.5:0.95 of 50.7%,outperforming the baseline by 8.6,9.1,8.7,and 9.0 percentage points,respectively. [Conclusions] The proposed method effectively improves the detection accuracy of personnel and protective equipment in railway UAV imagery,providing technical support for construction safety localization and railway safety management.

Key words: UAV image, UAV inspection, personal protective equipment detection, human spatial-structure prior, small object detection

中图分类号: