基于无监督图像增强的低光照车道线检测  

Lane Line Detection in Low Light Based on Unsupervised Image Enhancement

作　　者：赵玖龙 陈紫强[1] 姜弘岳 张骞 ZHAO Jiu-long;CHEN Zi-qiang;JIANG Hong-yue;ZHANG Qian(Guilin University of Electronic Technology,Guilin Guangxi 541004,China)

机构地区：[1]桂林电子科技大学,广西桂林541004

出　　处：《计算机仿真》2024年第9期116-120,共5页Computer Simulation

基　　金：国家自然科学基金资助项目(61861011);广西重大科技项目(No.AA17204093);ADAS中车辆道路环境感知技术研究(C21YJM00RX0T);2021年广西八桂学者科研补助(廖桂生)(C21RSC03XY02)。

摘　　要：现存的车道线检测方法在光照良好条件下具有较好的检测性能,但在低光照条件下性能急剧下降。针对低光照场景下车道线特征不明显导致不易检的问题,在车道线检测算法的预处理阶段加入融合空洞卷积的ZERO-DCE网络,提出了一种基于无监督图像增强网络的车道线检测方法。首先在ZERO-DCE网络中融入空洞卷积来提升目标信息的捕捉能力,结合车道线色彩属性采用两阶段图像融合方法提升低光照情况下车道线特征,然后利用UFAST网络进行车道线检测。在CULane数据集的进行性能分析测试,结果表明:文中算法相较于baseline算法,在正常光照环境下性能表现基本一致;在阴影环境和夜间环境下的F1值分别提升3.5和2.3。Existing lane detection methods have good detection performance under good illumination conditions,but their performance deteriorates sharply under low illumination conditions.Aiming at the problem that the features of lane lines are not obvious in low light scenes,this paper adds ZERO-DCE network with fusion of void convolution in the pre-processing stage of lane detection algorithm,and proposes a lane detection method based on unsupervised image enhancement network.Firstly,hollow convolution was integrated into the ZERO-DCE network to improve the capturing ability of target information.Combined with the color attributes of the lane lines,two-stage image fusion method was used to improve the lane line features under low light conditions,and then UFAST network was used to detect the lane lines.The performance analysis tests on CULane data set show that compared with baseline algorithm,the performance of the proposed algorithm was basically the same under normal lighting environment.F1 values in shadow and night environments increased by 3.5 and 2.3,respectively.

关 键 词：神经网络 图像增强 车道线检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]