基于改进RT-DETR的路面坑槽检测模型  

Pavement pothole detection model based on improved RT-DETR

作　　者：许小伟[1] 陈燕玲 占柳 漆庆华 邓明星[1] XU Xiaowei;CHEN Yanling;ZHAN Liu;QI Qinghua;DENG Mingxing(College of Automobile and Traffic Engineering,Wuhan University of Science and Technology,Wuhan 430065,China)

机构地区：[1]武汉科技大学汽车与交通工程学院,湖北武汉430065

出　　处：《武汉科技大学学报》2024年第6期457-467,共11页Journal of Wuhan University of Science and Technology

基　　金：国家重点研发计划项目(2022YFE0125200);国家自然科学基金项目(51975426);湖北省重点研发计划项目(2021BAA018,2022BAA062).

摘　　要：路面坑槽对驾驶的舒适性和安全性有很大影响。针对路面图像中坑槽尺寸小和特征信息匮乏导致检测精度低的问题,提出一种基于RT-DETR的路面坑槽检测模型Pavement Pothole-DETR(PP-DETR)。其主干网络使用SPDRSFE模块进行特征提取,可保留更多特征信息,提高小目标检测精度;引入渐进特征金字塔网络实现特征融合,避免多级传输造成的信息丢失,以解决坑槽特征信息主要集中在中、底特征层的问题;使用结构重参数化模块Conv3XCC3进行特征再提取,在提高网络表达能力的同时又不增加计算量。实验结果显示,相比原RT-DETR模型,PP-DETR的精确率与召回率分别提升了2.9和5.4个百分点,mAP达到76.9%。本文提出的改进方法有效提升了网络的特征提取和特征融合能力,在路面坑槽检测任务上的表现明显优于YOLO系列模型。Pavement potholes have a great impact on driving comfort and safety.To address the issue of low detection precision caused by small pothole size and lack of feature information in road surface images,a pothole detection model based on RT-DETR,named Pavement Pothole-DETR(PP-DETR),was proposed.PP-DETR took SPDRSFE module as the backbone network for feature extraction,which can retain more feature information and improve the detection precision for small targets.PP-DETR also introduced an asymptotic feature pyramid network for feature fusion,so as to avoid information loss due to multi-level transmission and solve the problem that the pothole feature information is mainly concentrated in the middle and bottom feature layers.By using structure re-parameterization module Conv3XCC3 for feature re-extraction,PP-DETR enhanced its network expressive ability without increasing the computational load.Experimental results show that,compared with the original RT-DETR model,PP-DETR’s precision and recall rate increase by 2.9 and 5.4 percentage points,respectively,and the mAP reaches 76.9%.The proposed method effectively improves the capability of feature extraction and feature fusion of the network,significantly outperforming the YOLO series models in pavement pothole detection tasks.

关 键 词：目标检测 路面坑槽 改进RT-DETR 渐进特征金字塔网络 结构重参数化 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]