基于深度生成网络的夜间车道线检测方法

doi:10.3969/j.issn.1674-8484.2025.03.011

摘要/Abstract

摘要：

为保障夜间车辆的安全行驶，准确识别夜间车道线并做出车道偏离预警，提出了用于夜间图像增强的深度生成网络EnhanceGAN和基于Transformer的端到端车道线检测网络AttentiveLSTR的夜间车道线检测方法，并进行实车实验。深度生成网络EnhanceGAN将改进后的UNet作为网络的生成器，采用两层嵌套的U形结构扩大感受野，添加Markov局部判别器和组合损失函数增强车道线边缘、纹理等细节信息；车道线检测网络AttentiveLSTR使用ResNeXt作为特征提取网络来保证网络深度和降低模型参数量，引入特征金字塔网络（FPN）提取车道线边缘和形状信息。结果表明：与主流方法CycleGAN和Gamma校正相比，该方法在BDD100k数据集上的夜间图像增强的效果更好，车道线和周围环境对比度高，结构相似性（SSIM）为0.8834，图像整体自然逼真，峰值信噪比（PSNR）为40.2654, 自然图像质量评估指标（NIQE）为3.4233；在CULane数据集上检测精度（Acc）为 90.12%，处理速度较快，每秒帧数（FPS）为82帧。该研究结果可以为夜间行驶车道线偏离场景提供参考。

关键词: 智能驾驶, 汽车安全, 生成网络, 夜间场景, 车道线检测, 深度学习

Abstract:

In order to ensure the safe driving of vehicles at night, the night lane lines were accurately recognized and lane departure warnings were made, a deep generative network EnhanceGAN for nighttime image enhancement and an end-to-end lane line detection network AttentiveLSTR based on Transformer were proposed for nighttime lane line detection, and experiments with real vehicles were conducted. The deep generative network EnhanceGAN used the improved UNet as the generator of the network, adopted a two-layer nested U-shape structure to expand the sensory field, and added a Markov local discriminator and a combined loss function to enhance the detailed information of lane line edges and textures. The lane line detection network AttentiveLSTR used ResNeXt as a feature extraction network to ensure the network depth and reduced the number of model parameters, and introduced feature pyramid networks (FPN) to extract lane line edge and shape information. The results show that compared with the mainstream methodsCycleGAN and Gamma Correction, the pro[osed method is more effective in nighttime image enhancement on the BDD100k dataset, with a high contrast between lane lines and surrounding environment, structural similarity (SSIM) of 0.883 4, natural and realistic images as a whole, peak signal-to-noise ratio (PSNR) of 40.265 4, and natural image quality evaluation index (NIQE) of 3.423 3; the detection accuracy (Acc) on the CULane dataset is 90.12%, and the processing speed is fast, with 82 frames per second (FPS). The research results can provide a reference for nighttime lane line deviation scenarios.

Key words: intelligent driving, automotive safety, generation network, night scenes, lane detection, deep learning

中图分类号:

TP391

刘国盛, 苏欣儿, 王建锋, 刘臻玮. 基于深度生成网络的夜间车道线检测方法[J]. 汽车安全与节能学报, 2025, 16(3): 452-462.

LIU Guosheng, SU Xiner, WANG Jianfeng, LIU Zhenwei. Night lane detection method based on deep generation network[J]. Journal of Automotive Safety and Energy, 2025, 16(3): 452-462.

图/表 15

参考文献 23

[1]	高轶男, 巩建强. 我国道路交通事故特征及致因分析[J]. 安全与环境学报, 2023, 23(11): 4013-4023.
	GAO Yinan, GONG Jianqiang. Characteristics and mechanism of road traffic accidents[J]. J Safe Environ, 2023, 23(11): 4013-4023. (in Chinese)
[2]	李福俊, 顾敏明. 车道线实时检测与偏离预警系统设计与研究[J]. 光电子激光, 2018, 29(3): 298-304.
	LI Fujun, GU Minming. Design and research on lane detection and the departure warning system[J]. J Optoelect Lase, 2018, 29(3): 298-304. (in Chinese)
[3]	Kang J, Yoon T, Kim E, et al. Lane-level map-matching method for vehicle localization using GPS and camera on a high-definition map[J]. Sensors, 2020, 20(8): 2166-2188.
[4]	Rastiveis H, Shams A, Sarasua W, et al. Automated extraction of lane markings from mobile LiDAR point clouds based on fuzzy inference[J]. ISPRS J Photogram Remo Sens, 2019, 160: 149-166.
[5]	Munir F, Azam S, Jeon M, et al. LDNet: End-to-end lane marking detection approach using a dynamic vision sensor[J]. IEEE Trans Intel Transport Syst, 2021, 23(7): 9318-9334.
[6]	Ding D, Lee C, Lee K-Y. An adaptive road ROI determination algorithm for lane detection[C]// 2013 IEEE Int’l Conf IEEE Regi 10 (TENCON 2013), Xi'an, China, 2013: 1-4.
[7]	FU Mengyin, WANG Xinyu, MA Hongbin, et al. Multi-lanes detection based on panoramic camera[C]// 11th IEEE Int’l Conf Contr Auto (ICCA), Taichung, Taiwan, 2014: 655-660.
[8]	陈洋, 黄孝慈, 吴训成. 基于改进卡尔曼滤波的车道线与车辆跟踪系统算法研究[J]. 计算机与数字工程, 2021, 49(7): 1363-1366.
	CHNEG Yang, HUANG Xiaoci, WU Xuncheng. Research on algorithm of lane line and vehicle tracking system based on improved Kalman filter[J]. Comput Digit Engi, 2021, 49(7): 1363-1366. (in Chinese)
[9]	Kim J-G, Yoo J-H, Koo J-C. Road and lane detection using stereo camera[C]// 2018 IEEE Int’l Conf Big Data Smart Comput (Big Comp), Shanghai, China, 2018: 649-652.
[10]	杨淑琴, 马玉浩, 方铭宇, 等. 基于实例分割的复杂环境车道线检测方法[J]. 浙江大学学报, 2022, 56(4): 809-815.
	YANG Shuqi, MA Yuhao, FANG Minyu, et al. Lane detection method in complex environments based on instance segmentation[J]. J Zhejiang University, 2022, 56(4): 809-815. (in Chinese)
[11]	YAO Yifan, XIONG Huilin. LaneFormer: An efficient transformer-based network for fast lane detection[C]// 2022 China Auto Congress (CAC), Xiamen, 2022: 3111-3116.
[12]	LI Xiang, LI Jun, HU Xiaolin, et al. Line-CNN: End-to-end traffic line detection with line proposal unit[J]. IEEE Trans Intel Transport Syst, 2019, 21(1): 1-11.
[13]	宋扬, 李竹. 基于深度图像增强的夜间车道线检测技术[J]. 计算机应用, 2019, 39(S2): 103-106.
	SONG Yang, LI Zhu. Lane detection method based on image enhancement technique under night condition[J]. J Comput Appl, 2019, 39(S2): 103-106. (in Chinese)
[14]	郭克友, 王艺伟, 郭晓丽. LDA 与 LSD 相结合的车道线分类检测算法[J]. 计算机工程与应用, 2017, 53(24): 219-225. doi: 10.3778/j.issn.1002-8331.1606-0116
	GUO Keyo, WANG Yiwei, GUO Xiaoli. Lane classification algorithm combined LDA and LSD[J]. Comput Engi Appl, 2017, 53(24): 219-225. (in Chinese)
[15]	HE Gaoming, WANG Mei. Research on lane detection algorithm based on wavelet analysis and hough transform[C]// 2022 IEEE 21st Int’l Conf Ubiquit Comput Commun (IUCC/CIT/DSCI/SmartCNS), Chongqing: IEEE Press, 2022: 280-287.
[16]	CHEN Yiman, XIANG Zhiyu, DU Wentao. Improving lane detection with adaptive homography prediction[J]. Visu Comput, 2023, 39(2): 585-591.
[17]	LI Shoubiao, WU Xin, WU Zhifei, et al. Efficient multi-lane detection based on large-kernel convolution and location[J]. IEEE Access, 2023, 11: 58125-58135.
[18]	LIU Ruijin, YUAN Zejian, LIU Tie, et al. End-to-end lane shape prediction with transformers[C]// 2021 IEEE Winter Conf Appl Comput Visi (WACV), Waikoloa: IEEE Press, 2021: 3693-3701.
[19]	GAO Rong, HU Siqi, YAN Lingyu, et al. High-order deep infomax-guided deformable transformer network for efficient lane detection[J]. Sign, Imag Vide Proc, 2023, 17(6): 3045-3052.
[20]	王杰, 陈黎卿, 黄莉莉, 等. 基于Retinex的弱光条件下车道线识别方法[J]. 计算机与数字工程, 2019, 47(2): 451-456.
	WANG Jie, CHEN Liqi, HUANG Lili. Lane Recognition method in weak light condition based on Retinex[J]. Comput Digit Engi, 2019, 47(2): 451-456. (in Chinese)
[21]	MA Hui, LÜ Wenhao, LI Yu, et al. Image adaptive contrast enhancement for low-illumination lane lines based on improved retinex and guided filter[J]. Appl Artifi Intel, 2021, 35(15): 1970-1989.
[22]	周经美, 王钰, 宁航, 等. 面向多元场景结合GLNet的车道线检测算法[J]. 中国公路学报, 2021, 34(7): 118-127. doi: 10.19721/j.cnki.1001-7372.2021.07.009
	ZHOU Jingmei, WANG Yu, NING Hang, et al. Lane detection algorithm based on GLNet for multiple scenes[J]. China J Highw Transport, 2021, 34(7): 118-127. (in Chinese)
[23]	Son J, Yoo H, Kim S, et al. Real-time illumination invariant lane detection for lane departure warning system[J]. Expet Syst Appl, 2015, 42(4): 1816-1824.

评价指标	PSNR	SSIM	NIQE	Acc / %	FP / %	FN / %
原始图像			4.152 2	83.72	21.96	25.09
只增加注意力模块	29.828 2	0.572 8	3.808 3	85.81	18.46	18.51
只增加U2Net生成器	30.342 1	0.791 6	3.632 1	86.31	33.01	19.02
EnhanceGAN	32.441 3	0.898 8	3.201 2	88.76	16.1 4	17.99

评价指标	PSNR	SSIM	NIQE	Acc / %	FP / %	FN / %
原始图像			4.152 2	83.72	21.96	25.09
只增加注意力模块	29.828 2	0.572 8	3.808 3	85.81	18.46	18.51
只增加U2Net生成器	30.342 1	0.791 6	3.632 1	86.31	33.01	19.02
EnhanceGAN	32.441 3	0.898 8	3.201 2	88.76	16.1 4	17.99

评价指标	PSNR	SSIM	NIQE	Acc / %	FP / %	FN / %
原始图像			4.453 1	83.72	21.96	25.09
CycleGAN	35.337 8	0.782 2	3.675 2	86.87	19.72	21.14
Gamma校正	34.219 8	0.690 3	3.821 5	87.21	18.89	19.94
EnhanceaGAN	40.265 4	0.883 4	3.423 3	88.76	16.14	17.99

评价指标	PSNR	SSIM	NIQE	Acc / %	FP / %	FN / %
原始图像			4.453 1	83.72	21.96	25.09
CycleGAN	35.337 8	0.782 2	3.675 2	86.87	19.72	21.14
Gamma校正	34.219 8	0.690 3	3.821 5	87.21	18.89	19.94
EnhanceaGAN	40.265 4	0.883 4	3.423 3	88.76	16.14	17.99

平台	模型	Acc %	FP %	FN %	Parameter	FPS (帧 · s^-1)
PC	LSTR	87.20	7.62	4.30	765 787	137
PC	本文	90.12	2.92	3.41	2 808 015	82
工控机	LSTR	86.36	9.42	8.67	765 787	78
工控机	本文	89.51	6.23	3.79	2 808 015	51