嵌入注意力机制的车道线像素级识别算法研究  

Pixel-level recognition algorithm of lane embedded with attention mechanism

作　　者：肖庭舒 罗小龙[1] 相龙伟 陈阳光 王朋燕 XIAO Tingshu;LUO Xiaolong;XIANG Longwei;CHEN Yangguang;WANG Pengyan(College of Earth Sciences,Yangtze University,Wuhan 430100,China)

机构地区：[1]长江大学地球科学学院,武汉430100

出　　处：《激光杂志》2025年第2期106-114,共9页Laser Journal

基　　金：国家自然科学基金(No.42004007)。

摘　　要：车辆自动行驶的安全性和稳定性离不开车道线准确识别。然而,日常驾驶中面临着复杂多变的天气和光照条件、道路标记模糊或遮挡等挑战。研究并设计基于深度神经网络的车道线识别算法,以提高识别技术在面对复杂环境的鲁棒性与检测结果精度。通过构建以VGG-16为主链并嵌入通道注意力和空间注意力机制的全卷积神经网络模型,实现端到端像素级别的车道线语义分割。嵌入注意力模块的新模型在CULane通用数据集上验证结果同VGG-解码语义分割方法相比,其平均像素准确率与均交并比(Mean Intersection over Union, MIoU)分别提升2.2%与1.3%。且在车道线不存在场景下,预测结果的像素准确率达到70%。嵌入注意力机制的图像分割算法研究为车道线识别问题提供了有效解决方案,有力支撑车道线检测技术在无人驾驶场景的应用。The safety and stability of autonomous vehicle driving are inseparable from accurate lane recognition.However,daily driving faces challenges such as complex and changing weather and lighting conditions,blurred or blocked road markings.Research and design lane line recognition algorithms based on deep neural networks to improve the robustness of recognition technology in complex environments and the accuracy of detection results.By constructing a fully convolutional neural network model with VGG-16 as the main chain and embedding channel attention and spa-tial attention mechanisms,end-to-end pixel-level lane lines semantic segmentation is achieved.The new model em-bedding the attention module is verified on the CULane general data set.Compared with the VGG-decoding semantic segmentation method,its average pixel accuracy and Mean Intersection over Union(MIoU)increased by 2.2%and 1.3%respectively.And in the scenario where lanes do not exist,the pixel accuracy of the prediction results reaches 70%.Research on image segmentation algorithms embedding attention mechanisms provides an effective solution to the problem of lane line recognition,and strongly supports the application of lane line detection technology in driverless driving scenarios.

关 键 词：注意力机制 深度神经网络 语义分割 车道线识别 图像分割 

分 类 号：TN209[电子电信—物理电子学]