| 张全,周甯,刘洋毅,段昶,李艳,彭博.基于特征融合与行锚点分类的车道线快速检测算法[J].电子测量与仪器学报,2025,39(12):188-196 |
| 基于特征融合与行锚点分类的车道线快速检测算法 |
| Fast lane detection algorithm based on feature fusion androw anchor classification |
| |
| DOI： |
| 中文关键词: 车道线检测 多尺度特征 卷积神经网络 ASPP 特征金字塔网络 |
| 英文关键词:lane detection multi-scale features convolutional neural network ASPP feature pyramid network |
| 基金项目:智能警务四川省重点实验室开放课题(ZNJW2024KFMS003)项目资助 |
|
|
| 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, Sichuan Police College, Luzhou 646000, China |
| Duan Chang | School of Electrical Engineering and Information, Southwest Petroleum University, Chengdu 610500, China |
| Li Yan | School of Computer Science and Software Engineering, Southwest Petroleum University, Chengdu 610500, China |
| Peng Bo | School of Computer Science and Software Engineering, Southwest Petroleum University, Chengdu 610500, China |
|
| 摘要点击次数: 585 |
| 全文下载次数: 376 |
| 中文摘要: |
| 为了解决传统图像处理方法在阴影、夜晚等复杂场景下车道线检测实时性和准确性难以兼顾的问题，提出了一种基于特征融合与行锚点分类的车道线快速检测算法，以适应实时交通场景的需求。在图像预处理阶段，将图像划分为网格状的行锚点，将车道线检测转化为行锚点分类问题，显著降低计算量。车道线检测网络采用ResNet-18作为骨干网络，并引入聚合模块增强上下文特征提取，提升车道线结构信息捕捉能力；结合特征金字塔网络（FPN）进行多尺度特征融合，实现车道线局部特征与全局特征的互补；此外，引入含ASPP模块的辅助分割分支，进一步优化车道线检测精度。在公开数据集TuSimple和CULane上进行实验，TuSimple数据集上的精确度达到96.16%，运行耗时仅为3.2 ms；CULane数据集上取得了70.3%的F1分数，帧率达到310 fps。实验结果表明，所提方法在保证检测精度的同时，显著提高了检测速度。 |
| 英文摘要: |
| In order to solve the problem of difficulty in balancing real-time and accuracy of lane detection in complex scenes such as shadows and nights using traditional image processing methods, a fast lane detection algorithm based on feature fusion and anchor point classification is proposed to meet the needs of real-time traffic scenes. In the image preprocessing stage, the image is divided into grid like row anchors, and lane detection is transformed into a row anchor classification problem, significantly reducing computational complexity. The lane detection network adopts ResNet-18 as the backbone network and introduces an aggregation module to enhance context feature extraction and improve the ability to capture lane structure information. Combining feature pyramid network (FPN) to achieve multi-scale feature fusion and complement local and global features of lane markings. In addition, an auxiliary segmentation branch with ASPP module is introduced to further optimize the accuracy of lane detection. Experiments were conducted on the public datasets TuSimple and CULane, and the accuracy on the TuSimple dataset reached 96.16%, with a running time of only 3.2 ms; Obtained 70.3% F1 score and FPS of 310 fps on the CULane dataset. The experimental results show that the proposed method significantly improves detection speed while ensuring detection accuracy. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
