复杂环境下基于实例分割的车道线检测  

Lane Detection Based on Instance Segmentation in Complex Environment

作　　者：谢春丽 梁梓涵 XIE Chunli;LIANG Zihan(College of Mechanical and Electrical Engineering,Northeast Forestry University,Harbin 150040,Heilongjiang,China)

机构地区：[1]东北林业大学机电工程学院,黑龙江哈尔滨150040

出　　处：《重庆交通大学学报(自然科学版)》2025年第4期79-86,120,共9页Journal of Chongqing Jiaotong University(Natural Science)

基　　金：黑龙江省自然科学基金项目(LH2021F002)。

摘　　要：面对复杂场景下车道线检测存在适应性不够,检测精度不足的问题,笔者提出了一种基于实例分割的车道检测与跟踪算法。该算法利用Encoder-Decoder网络模型来执行实例划分,并引入双注意力机制模块(convolutional block attention module,CBAM)提升车道线分割的准确性,获取能有效隔离道路特性与环境干扰的划分图像;最后结合可变透视变换矩阵变换道路图像,采用多项式拟合生成车道线参数方程。研究结果显示:该算法在TuSimple数据集上的检测精度达到96.60%,在CULane数据集上F_(S1)分数达到79.8%;不仅显著提升车道线的分割速度,且在复杂交通环境中有良好的鲁棒性和检测精度。Aiming at the problem of insufficient adaptability and detection accuracy in lane detection in complex scenes,a lane detection and tracking algorithm based on instance segmentation was proposed.The proposed algorithm employed an Encoder-Decoder network model for instance segmentation and incorporated a convolutional block attention module(CBAM)to enhance the accuracy of lane line segmentation.The segmented images that could effectively isolate road characteristics from environmental interference were obtained.Finally,combining the variable perspective transformation matrix to transform the road image,polynomial fitting was used to generate the lane line parameter equation.The research results show that the proposed algorithm achieves a detection accuracy of 96.60%on the TuSimple dataset and an F_(S1) score of 79.8%on the CULane dataset,which not only significantly improves the segmentation speed of lane lines but also has good robustness and detection accuracy in complex traffic environments.

关 键 词：交通运输工程 车道线检测 实例分割 注意力机制 透视变换 

分 类 号：U463.6[机械工程—车辆工程]