| 张全,周甯,刘洋毅,段昶.先验引导的多尺度车道线检测网络[J].电子测量与仪器学报,2026,40(4):236-244 |
| 先验引导的多尺度车道线检测网络 |
| Multi scale lane detection network guided by prior guidance |
| |
| DOI： |
| 中文关键词: 车道线检测 车道先验 注意力机制 特征融合 交叉熵损失 |
| 英文关键词:lane detection lane prior attention mechanism feature fusion cross entropy loss |
| 基金项目:智能警务四川省重点实验室开放课题(ZNJW2024KFMS003)、泸州市应用基础研究（2024JYJ055)资助项目 |
|
|
| Author | Institution |
| Zhang Quan | School of Computer Science and Software Engineering, Southwest Petroleum University, Chengdu 610500, China |
| Zhou Ning | School of Computer Science and Software Engineering, Southwest Petroleum University, Chengdu 610500, China |
| Liu Yangyi | Intelligent Policing Key Laboratory of Sichuan Province, Luzhou 646000, China |
| Duan Chang | School of Electrical Engineering
and Information, Southwest Petroleum University, Chengdu 610500, China |
|
| 摘要点击次数: 285 |
| 全文下载次数: 123 |
| 中文摘要: |
| 精准的车道线信息对于自动驾驶系统中的路径规划和决策模块至关重要。为实现高精度的车道线检测，将车道线细长且结构连续的特点作为先验信息，提出一种先验引导的多尺度车道线检测网络。该网络以Resnet-18作为主干特征提取结构，在不同阶段引入门控融合模块，有效抑制了冗余信息传播，提高了跨层信息交互效率。针对车道线细长、连续、跨越大范围空间的特点，在FPN网络中引入了大核注意力机制，显著扩大感受野，增强网络对不同尺度车道线的感知能力。在检测头部分，结合了车道先验和全局特征，有效解决了仅依赖局部像素难以准确识别车道线的问题。此外，设计辅助分割分支并采用交叉熵损失函数，以优化参数更新过程，增强模型在细粒度特征上的表达能力。实验表明，在TuSimple数据集上精确度达到96.96%；CULane数据集上取得了80.1%的F1分数，该方法在复杂场景能够实现了高精度的车道线检测，且在现有的车道线检测方法中展现出很强的竞争力。 |
| 英文摘要: |
| Accurate lane line information is crucial for path planning and decision-making modules in autonomous driving systems. To achieve high-precision lane detection, this paper leverages the inherent characteristics of lane lines—being slender and structurally continuous—as prior knowledge, proposing a lane-prior guided multi-scale lane detection network. The network employs ResNet-18 as the backbone feature extraction structure. Gated fusion modules are introduced at different stages to effectively suppress redundant information propagation and enhance cross-layer information interaction efficiency. Addressing the slender, continuous, and spatially extensive nature of lane lines, a large kernel attention mechanism is incorporated into the feature pyramid network (FPN), significantly expanding the receptive field and strengthening the network’s perception capability for lane lines across different scales. In the detection head, lane priors are combined with global features, effectively resolving the challenge of accurately identifying lane lines when relying solely on local pixels. Furthermore, an auxiliary segmentation branch is designed, utilizing a cross-entropy loss function to optimize the parameter update process and enhance the model’s expressive power for fine-grained features. Experiments conducted on the public TuSimple and CULane datasets demonstrate the effectiveness of the proposed method: it achieves 96.96% accuracy on the TuSimple dataset and an F1-score of 80.1% on the CULane dataset. The proposed method enables high-precision lane detection in complex scenarios and exhibits strong competitiveness among existing lane detection approaches. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
