基于改进LaneNet网络的车道线检测方法  

Lane line detection method based on improved LaneNet network

作　　者：刘宏建 张伟斌 LIU Hongjian;ZHANG Weibin(School of Electronic and Optical Engineering,Nanjing University of Science and Technology,Nanjing 210014,China)

机构地区：[1]南京理工大学电子工程与光电技术学院,江苏南京210014

出　　处：《现代交通与冶金材料》2025年第2期39-50,共12页Modern Transportation and Metallurgical Materials

基　　金：国家自然科学基金资助项目(71971116)。

摘　　要：科技发展日新月异,自动驾驶成为大家研究的热门领域。受复杂环境的影响,车道线检测很容易出现漏检误检的情况。传统的LaneNet车道线检测算法可以通过像素级别的图像处理来检测车道线,但该算法没有区分图像的关键区域,所以在复杂环境中,其检测能力大幅下降。为了提高其检测能力,本文对LaneNet网络进行了优化。提出了一种引入边缘特征和U-Net网络的语义分割模型Edge-Feature U-Net LaneNet(简称EU-LaneNet),该模型更加关注车道线附近像素的变化,还能一定程度宽容车道线的形变,从而使模型具有更好的鲁棒性。EU-LaneNet模型中使用了UNet网络作为编解码结构,U-Net网络将编码器与解码器进行跳跃层连接,从而保留了更多的空间信息和上下文信息,这有助于保留更多的细节并提高分割结果的准确性。在EU-LaneNet模型中增加了空洞空间卷积池化金字塔(ASPP, Atrous Spatial Pyramid Pooling)与频率域通道注意力(FCANet, Frequency Channel Attention)的融合机制,该方法具备从广阔的感受野中捕捉丰富的上下文信息的能力,同时提取出便于分析的细节特征,并通过这些细节特征来抑制噪声。利用自己构建的道路监控数据集和Tusimle车道线检测数据集对该模型进行综合训练。车道线检测结果显示,本文提出的模型效果更好,在准确率基本保持不变的情况下,本研究提出的车道线检测综合模型相对于传统的车道线检测LaneNet网络精确率提高了8.96%,能够更好地适应复杂的环境。With the rapid development of science and technology,automatic driving has become a hot research field.In the process of lane line detection,affected by the complex environment,lane line de-tection is prone to miss detection and false detection.How to accurately identify lane lines has become a difficult point in automatic driving research.The traditional LaneNet lane line detection algorithm can detect and locate lane lines through pixel level image processing,but this algorithm does not distin-guish the key areas of the image,so its detection ability drops significantly in complex environments.In order to improve its detection ability,this paper optimizes the LaneNet network.A semantic segmenta-tion model,Edge Feature U-Net LaneNet(EU LaneNet)is proposed,which introduces edge features and U-Net networks.This model pays more attention to the changes in pixels near the lane line and can tolerate the deformation of the lane line to a certain extent,making the model more robust.The EU LaneNet model uses the U-Net network as the encoding and decoding structure.The U-Net network connects the encoder and decoder through skip layers,thereby retaining more spatial and contextual in-formation,which helps to retain more details and improve the accuracy of segmentation results.A fusion mechanism of Atrous Spatial Pyramid Pooling(ASPP)and Frequency Channel Attention(FCANet)is added to the EU LaneNet model.This method has the ability to capture rich contextual information from a wide receptive field,extract detailed features that are easy to analyze,and suppress noise through these detailed features.The model is comprehensively trained by using the road monitoring dataset built by myself and the Tusimle lane line detection dataset.The results of lane line detection show that the model proposed in this paper performs better.While the accuracy remains largely un-changed,the comprehensive model for lane line detection proposed in this study improves the accuracy by 8.96%compared to the traditional LaneNet network for lane line

关 键 词：智能交通 车道线检测 LaneNet U-Net 空洞空间卷积池化金字塔 频率域通道注意力 

分 类 号：U491[交通运输工程—交通运输规划与管理]