摘要
针对道路转弯处曲率过大导致弯道识别的精度下降的问题,本文提出了一种基于循环特征融合Resa-CC的弯道增强车道线检测算法。该算法利用车道线的形状先验性,捕获图像像素中行与列的空间关系,融合信息生成特征图;以残差网络为主体框架,加入编码器、解码器和注意力机制模块,在损失函数中引入弯道结构约束来提高车道线弯道的识别精度。加入循环特征融合模块和自注意力机制模块后准确率分别提升3.41%和1.1%,证明了两模块的有效性;Resa-CC算法准确率可达96.83%,FPS为35.68,误检率FP和漏检率FN分别为0.0315和0.0282,表明本文算法具有较高的检测性能,在车辆行驶弯道路段中能更准确地推断出车道线的位置。
A curve enhanced lane detection algorithm based on cyclic feature fusion Resa-CC is proposed to address the issue of reduced accuracy in curve recognition caused by excessive curvature at road turns.This algorithm utilizes the shape priors of lane lines to capture the spatial relationships between rows and columns in image pixels,and fuse information to generate feature maps.The residual network is used as the main framework,and the encoder,decoder and attention mechanism modules are added.The Loss function adds curve structure constraints to improve the recognition accuracy of lane curves.The addition of the cyclic feature fusion module and the self attention mechanism module improved the accuracy by 3.41%and 1.1%,respectively,proving the effectiveness of the two modules.The accuracy of the Resa-CC algorithm can reach 96.83%,with an FPS of 35.68.The false detection rate FP and missed detection rate FN are 0.0315 and 0.0282,respectively.This indicates that the algorithm has high detection performance and can more accurately infer the position of the lane line in the curve when vehicles are driving.
作者
鲁维佳
刘泽帅
潘玉恒
李国燕
李慧洁
丛佳
LU Weijia;LIU Zeshuai;PAN Yuheng;LI Guoyan;LI Huijie;CONG Jia(Tianjin Chengjian University,School of Computer Information and Engineering,Tianjin 300384,China;Tianjin Branch Company of China Tower Corperation Limited,Tianjin 300011,China)
出处
《测绘通报》
CSCD
北大核心
2023年第12期25-30,共6页
Bulletin of Surveying and Mapping
基金
国家自然科学基金(62204168)
天津市科技计划(20YDTPJC00160,21YDTPJC00780)
天津市教委科研计划(2019KJ101)。
关键词
交通工程
车道线检测
循环特征聚合
深度学习
自动驾驶
ResNet
traffic engineering
lane line detection
recurrent feature aggregation
deep learning
automatic driving
ResNet
作者简介
鲁维佳(1980-),男,硕士,讲师,研究方向为嵌入式系统设计、自动驾驶。E-mail:13920156816@139.com;通讯作者:潘玉恒,E-mail:panyuheng@tju.edu.cn。
