基于明度动态感知裁剪和YOLOv8的铁路沿线异常感知检测方法  

Anomaly Perception and Detection Method Along Railway Lines Based on Dynamic Value Perception and Image Cropping Along with YOLOv8

作　　者：林俊亭[1] 陈权 马赫 彭嘉维 柴金川 Lin Junting;Chen Quan;Ma He;Peng Jiawei;Chai Jinchuan(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;School of Information and Intelligent Manufacturing,Chongqing City Vocational College,Chongqing 402160,China;Changchun Signalling and Communication Depot,China Railway Shenyang Group Co.,Ltd.,Changchun 116019,China;National Railway Track Test Center,China Academy of Railway Sciences Corporation Limited,Beijing 100015,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,兰州730070 [2]重庆城市职业学院信息与智能制造学院,重庆402160 [3]中国铁路沈阳局集团有限公司长春电务段,长春116019 [4]中国铁道科学研究院集团有限公司国家铁道试验中心,北京100015

出　　处：《铁路通信信号工程技术》2025年第1期19-27,共9页Railway Signalling & Communication Engineering

基　　金：国家自然科学基金地区项目(No.52162050)。

摘　　要：铁路轨道沿线场景具有空旷开阔、占地面积大的特点。由于高清摄像头的普及,导致大部分入侵物体目标在图像中具有占比面积小,像素点数量少,进而降低算法对目标尤其是小目标的检测性能。针对该问题,提出基于明度动态感知裁剪和YOLOv8的铁路沿线异常感知检测方法。首先通过明度动态感知方法感知运动目标区域,再以动态目标为中心裁剪图片输入YOLOv8网络进行检测,防止目标区域图像输入网络时特征信息被压缩,最后在RailD49数据集上验证了该算法的召回率为80.33%,准确率为81.67%,相较于原始YOLOv8m对铁路运动目标的检测能力有所提升。Since the areas along the railway tracks are open and expansive,and as high-definition cameras are widely used,most of the intruding objects have the characteristics of occupying a small space in the images and exhibiting low pixel density,which reduces the detection performance of existing algorithms for targets,especially small ones.To address this problem,this paper proposes an anomaly perception and detection method along the railways based on dynamic value perception and image cropping,along with YOLOv8.First,the moving target area is perceived by the method of dynamic value perception.Then,the image is cropped with the dynamic target at the center and input into the YOLOv8 network for detection,to prevent the compression of the information about the key features when the target area image is fed into the network.Finally,the proposed algorithm is verified on the RailD49 dataset,with a recall of 80.33%and a precision of 81.67%,and is found to have better detection capability for moving targets along railways than original YOLOv8m.

关 键 词：目标检测 铁路综合视频监控系统 明度动态感知 深度学习 YOLOv8 

分 类 号：U284[交通运输工程—交通信息工程及控制] TP391.41[交通运输工程—道路与铁道工程]