改进YOLOv5s的铁路异物入侵检测算法  

Improved Railroad Foreign Object Intrusion Detection Algorithm for YOLOv5s

作　　者：孟彩霞 王兆楠[2] 石磊 高宇飞[2] 卫琳[2] MENG Caixia;WANG Zhaonan;SHI Lei;GAO Yufei;WEI Lin(School of Computer and Artificial Intelligence,Zhengzhou University,Zhengzhou 450001,China;School of Cyber Science and Engineering,Zhengzhou University,Zhengzhou 450002,China;Department of Image and Network Investigation Technology,Railway Police College,Zhengzhou 450053,China)

机构地区：[1]郑州大学计算机与人工智能学院,郑州450001 [2]郑州大学网络空间安全学院,郑州450002 [3]铁道警察学院图像与网络侦查系,郑州450053

出　　处：《小型微型计算机系统》2024年第4期879-886,共8页Journal of Chinese Computer Systems

基　　金：2022年度河南省重大科技专项(221100210100)资助;2020年度河南省重大公益专项项目(201300210500)资助;郑州大学高层次人才科研启动基金项目(32340306)资助.

摘　　要：行人、车辆等异物侵入铁路界线,严重威胁行人安全和铁路交通安全.针对传统铁路异物入侵检测方法精度低、时效性差等问题,提出改进YOLOv5s算法的铁路异物入侵检测模型SD-YOLO.本文提出SSA混合注意力机制,加强模型的局部表征能力,提高小目标识别效果;提出DW-Decoupled Head解耦检测头,加快网络收敛速度;引入边界框回归损失函数SIoU,提高了模型的检测精度;使用转置卷积作为采样方法,采样更适合铁路侵限障碍物特征的尺寸和比例.在数据集RS和Pascal VOC 2012进行实验验证,与基线YOLOv5s算法相比,平均精度mAP@0.5分别提高了2.7%、1.8%,mAP@.5:.95分别提高了2.9%、2.1%,检测速度分别达到79 FPS和78 FPS,表明该算法在检测精度和速度上均取得良好的性能,有效改善了漏检、误检问题,提高了小目标识别能力.The intrusion of foreign objects such as pedestrians and vehicles into railroad boundaries seriously threatens pedestrian safety and railroad traffic safety.To address the problems of low accuracy and poor timeliness in traditional railroad foreign object intrusion detection methods,a foreign object intrusion detection model SD-YOLO is proposed based on YOLOv5s improvement.This paper proposes SSA hybrid attention mechanism to enhance the local characterization ability of the model and improve the effect of small target recognition;proposes DW-Decoupled Head to decouple the detection head and speed up the network convergence;introduces Bounding box regression loss function SIoU to improve the detection accuracy of the model;using transposed convolution as the sampling method,the sampling is more suitable for the size and scale of the railroad encroachment limit obstacle features.Experimental verification was carried out in the dataset RS and Pascal VOC 2012.Compared with the baseline YOLOv5s algorithm,mAP@0.5 increased by 2.7%and 1.8%respectively,mAP@.5:.95 increased by 2.9%and 2.1%respectively,and the detection speed reached 79 FPS and 78 FPS respectively,indicating that the algorithm has achieved good performance in detection accuracy and speed,effectively improved the problems of missed detection and false detection,and improved the ability of small target recognition.

关 键 词：铁路入侵检测 混合注意力机制 解耦头 损失函数 上采样 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]