计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (13): 178-185.DOI: 10.3778/j.issn.1002-8331.2212-0283

• 图形图像处理 • 上一篇    下一篇

复杂场景中的快速车道线检测方法

杜黎,吕毅斌,武德安,罗雨欣   

  1. 1.昆明理工大学 理学院,昆明 650500
    2.电子科技大学 数学科学学院,成都 611731
  • 出版日期:2023-07-01 发布日期:2023-07-01

Fast Lane Detection Method in Complex Scenarios

DU Li, LYU Yibin, WU Dean, LUO Yuxin   

  1. 1.Faculty of Science, Kunming University of Science and Technology, Kunming 650600, China
    2.School of Mathematical Sciences, University of Electronic Science and Technology of China, Chengdu 611731, China
  • Online:2023-07-01 Published:2023-07-01

摘要: 车道线检测是实现无人驾驶系统的一项重要任务。针对当前基于分割的检测方法实时性能不足的问题,提出了一种快速车道线检测方法FracLane。设计一种高效的特征提取模块,通过分形的思想构造基于聚合不同大小感受野的分形残差结构,更准确地提取车道特征。结合龙格-库塔法,使用可训练参数对分形残差结构的输出进行加权融合,构造能进一步提升网络性能的分形残差模块(FracRes)。在特征解码阶段引入一种基于行锚检测方法的车道位置预测模块(LPP),极大提高网络的检测速度。在包含多种交通场景的车道检测数据集TuSimple和CULane上进行的大量实验,结果表明,该方法在两个数据集上最高可获得97.26%的准确率和78.1的F1评分,在800×288的分辨率下,最高可获得206?FPS的推理速度。与现有检测方法相比,该方法在检测精度与速度方面都有明显提高,达到实时检测任务的需求。

关键词: 车道线检测, 行锚, 分形残差, 龙格-库塔, 卷积神经网络

Abstract: Lane detection is an essential task of unmanned systems. This paper proposes a fast lane detection method FracLane to deal with the problem of the existing segmentation-based lane detection methods with insufficient real-time performance. Firstly, an efficient fractal residual structure based on receptive fields of different sizes is designed, which improves the lane feature extraction capability by introducing the idea of the Fractal. Then, combined with the Runge-Kutta method, the output of the fractal residual structure is weighted and fused using trainable parameters to construct a fractal residual module(FracRes) that can better improve the performance of the network. Finally, a lane position prediction module(LPP) based on row anchor detection is introduced into the decoder to improve the inference speed. A large number of experiments are conducted on TuSimple and CULane that consist of multiple traffic scenarios. Results show that the accuracy and F1-measure of the proposed method are 97.26% and 78.1, respectively, and the inference speed can reach 206?FPS. Compared with the existing methods, the proposed method shows significant improvement, which adequately meets real-time task requirements.

Key words: lane detection, row anchor, fractal residual, Runge-Kutta, convolutional neural network