基于图像网格化分割的实时车道线识别算法  

Real-time Lane Recognition Algorithm Based on Image Grid Segmentation

作　　者：谢荀 揭豪 洪伟荣[1] Xie Xun;Jie Hao;Hong Weirong

机构地区：[1]浙江大学能源工程学院,浙江省杭州市310027

出　　处：《工程机械》2021年第10期1-6,I0021,共7页Construction Machinery and Equipment

摘　　要：车道线识别技术是利用光学传感器对环境进行感知,为车辆行为的决策和控制提供重要的车道信息,其本质是从图像中提取车道线的轨迹坐标。基于深度学习的车道线识别算法大多采用密集预测型方法提取车道线轨迹,因而在提取过程会重复处理大量的冗余信息。针对这一问题提出一种改进算法,该算法将图像划分为多个小网格,并对各网格进行分类,从而减少提取过程中冗余信息的处理次数。在此基础上,设计深度神经网络用于网格分类,其与常规密集预测型神经网络的区别在于:利用了无参数的亚像素卷积替代转置卷积以避免棋盘格效应;采用1×1卷积替代全连接层分类器,实现在网格分类过程中保留关键空间位置信息,同时降低网络规模。在CULane数据集上分别训练Lane Net与改进算法,并进行对比分析,分析结果表明,改进后的车道线检测算法在准确性与实时性方面均具有明显优势。Lane recognition technology is to perceive the environment by optical sensors,providing important lane information for vehicle decision-making and control,of which the essence is to extract trajectory coordinates of lane from the image.Most lane recognition algorithms based on deep learning use dense prediction method to extract lane trajectories,resulting in repeatedly processing of a large amount of redundant information in the extraction process.To solve this problem,an improved algorithm is proposed,which divides the image into several small grids and classifies the grids to reduce the processing times of redundant information in the extraction process.On this basis,the deep neural network is designed for grid classification,which is distinguished from conventional dense prediction neural network in:use of non-parametric sub-pixel convolution instead of transpose convolution to avoid Checkerboard Effect;use of 1×1 convolution instead of full connection layer classifier to preserve key spatial location information in the process of grid classification while reducing network scale.LaneNet and the improved algorithm are trained on CULane data set respectively,compared and analyzed.The analysis results show that the improved lane detection algorithm has obvious advantages in accuracy and instantaneity.

关 键 词：网格化分割 车道线识别 深度学习 神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] U463.6[自动化与计算机技术—计算机科学与技术]