电子、通信与自动控制技术

结合增强空间感知的远距离车道线检测方法

展开
  • 兰州交通大学电子与信息工程学院, 甘肃 兰州 730070



网络出版日期: 2025-09-25

A Long-Range Lane Detection Method Integrating Enhanced Spatial Perception

Expand
  • School of Electronic and Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China

Online published: 2025-09-25

摘要

车道线检测作为智能汽车视觉导航系统的核心技术,直接影响到车辆的路径引导与转向控制,对提升交通安全性和导航效率具有重要意义。车道线图像中背景信息往往占据主导地位,特别是远距离车道线存在特征小、标记缺失或被遮挡等问题,同时伴随视觉上的宽度变化,导致远距离车道线比正常车道线更加难以识别。为了解决这一问题,本文提出了结合增强空间感知的车道线检测方法。针对车道线在图像中呈细长结构的特点,在主干网络中引入条形池化,以细化车道线信息的表示。提出增强空间感知优化器(ESAO)和车道线多尺度聚合器(LMSA)。ESAO增强小目标特征,有效区分目标和背景信息。LMSA是一种车道线形状感知多尺度聚合模块,能够增强每个锚点的局部特征表示与相关性,充分融合全局和局部特征信息,结合车道线的形状先验来完整地重构出整条车道线形状。此外,通过全局和局部斜率一致性损失函数,来自适应调整车道线形状和位置。实验结果表明,本文方法优于对比实验中最优秀的方法,该方法在TuSimple数据集上的F1分数和准确率分别提高0.58%和0.44%,在CULane数据集上的F1@50得分提高1.14%,特别是在远距离道路场景中表现的性能更加稳定。

本文引用格式

王耀琦, 卢亚琦, 王小鹏 . 结合增强空间感知的远距离车道线检测方法[J]. 华南理工大学学报(自然科学版), 0 : 1 . DOI: 10.12141/j.issn.1000-565X.250181

Abstract

As a core technology in the visual navigation systems of intelligent vehicles, lane detection directly influences path planning and steering control, playing a crucial role in improving both traffic safety and navigation efficiency. However, lane images are often dominated by background information, particularly in long-range scenarios where lane exhibit weak features, missing markings, or occlusions. These issues, compounded by perspective-induced width variations, make long-range lane significantly more difficult to detect than nearby ones.To address these challenges, this paper proposes a lane detection method that incorporates enhanced spatial perception. First, considering the elongated structural characteristics of lane in images, strip pooling is introduced into the backbone network to refine the representation of lane features. Second, we design an Enhanced Spatial-Aware Optimizer (ESAO) and a Lane Multi-Scale Aggregator (LMSA). The ESAO enhances small-scale feature representations and effectively distinguishes targets from background information. The LMSA is a lane-shape-aware multi-scale aggregation module that strengthens the local feature representation and correlation at each anchor point. It enables effective fusion of global and local features and reconstructs the complete lane shape by incorporating prior knowledge of lane geometry.In addition, a global and local slope consistency loss function is introduced to adaptively refine the shape and position of lanes.Experimental results demonstrate that the proposed method outperforms the best baseline approaches, achieving improvements of 0.58% in F1-score and 0.44% in accuracy on the TuSimple dataset, and a 1.14% increase in F1@50-score on the CULane dataset. Notably, the method exhibits more stable performance in long-distance road scenarios.
Options
文章导航

/