基于轻量化Transformer的车道线检测方法  

Lane line detection method based on lightweight transformer

作　　者：陈广秋[1] 刘枫铭 段锦[1] 黄丹丹 CHEN Guangqiu;LIU Fengming;DUAN Jin;HUANG Dandan(School of Electronic and Information Engineering,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学电子信息工程学院,吉林长春130022

出　　处：《华中科技大学学报(自然科学版)》2025年第3期117-126,共10页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62127813)。

摘　　要：自动驾驶和高级驾驶辅助系统在移动设备上部署时,由于网络参数量过多导致存储空间占用过大,硬件系统的部署门槛过高,影响自动驾驶和辅助驾驶技术的普及.为了解决上述问题,在语义分割网络的框架下,提出了一种基于轻量化Transformer车道线检测方法.在编码器部分,采用针对Transformer模块轻量化设计的MobileVIT网络感知全局依赖关系,捕获更远距离的车道线相关特征信息,降低网络参数量;在解码器部分,采用双边上采样解码器对分割结果进行精细化处理,得到更精确的像素级分割结果;最后利用置信度评估网络确定出车道线数量;此外,在网络训练阶段引入自注意蒸馏方法,在不增加网络参数量的同时,提高车道线区域的注意力.为了满足不同应用需求,设计了3个不同参数量的检测网络.实验结果表明:设计的3个网络,参数量为典型车道线检测网络SCNN-ResNet34的26.03%,13.19%和7.52%,准确率分别提高了0.46%,0.15%和0.09%,实现了在较少参数量的情况下,具有较高的检测准确率,便于在移动设备上部署.When deploying autonomous driving and advanced driver assistance systems on mobile devices,excessive network parameters resulted in large storage space occupation and high deployment threshold of hardware systems,affecting the popularization of autonomous driving and assisted driving technologies.To address these issues,a lane detection method based on lightweight transformer within the framework of a semantic segmentation network was proposed.In the encoder part,a MobileVIT network tailored for lightweight design of transformer modules was utilized to perceive global dependency relationships,capturing lane-related feature information at longer distances and reducing network parameter count.In the decoder part,a bilateral upsampling decoder was employed to refine the segmentation results,yielding more accurate pixel-level segmentation results.Finally,a confidence evaluation network was used to determine the number of lane lines.Additionally,a self-attention distillation method was introduced during network training to enhance the attention on lane line areas without increasing network parameters.To meet various application requirements,three detection networks with different parameter counts were designed.Experimental results demonstrate that the parameter counts of the three designed networks are 26.03%,13.19%,and 7.52% of the typical lane detection network SCNN-ResNet34,respectively.The accuracies are improved by 0.46%,0.15%,and 0.09% respectively,achieving high detection accuracy with fewer parameters,making it convenient for deployment on mobile devices.

关 键 词：交通工程 语义分割 车道线检测 MobileViT网络 自注意力蒸馏 

分 类 号：U463.6[机械工程—车辆工程] TP391.41[交通运输工程—载运工具运用工程]