计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (14): 227-235.DOI: 10.3778/j.issn.1002-8331.2012-0138

• 图形图像处理 • 上一篇    下一篇

雾天车道线识别方法:FoggyCULane数据集的创建

徐哲钧,张暐,郭昊,张洋,李庆,董雪   

  1. 上海交通大学 中英国际低碳学院,上海 200240
  • 出版日期:2022-07-15 发布日期:2022-07-15

Lane Detection Under Foggy Conditions via FoggyCULane Dataset

XU Zhejun, ZHANG Wei, GUO Hao, ZHANG Yang, LI Qing, DONG Xue   

  1. China-UK Low Carbon College, Shanghai Jiao Tong University, Shanghai 200240, China
  • Online:2022-07-15 Published:2022-07-15

摘要: 为了提高深度学习算法在雾天场景下的车道线识别率,扩充雾天车道线数据集是有效途径之一。以目前最具有权威性的CULane数据集为基础,通过对该数据集内晴天车道线图片进行单幅图像深度提取,随后依照大气散射模型生成3种不同浓度的雾天车道线图片,并保留原图的车道线标签,以此方法实现对CULane数据集的人为扩充。通过增加了107?451张带标签的雾天车道线图像,从而将原始CULane数据集扩充了1.8倍,建立了包含雾天图像的新车道线数据集FoggyCULane。分别采用原始CULane数据集和FoggyCULane数据集对SCNN车道识别网络进行训练,并将训练结果在包含3种不同浓度雾天场景的12种复杂车道线场景中进行测试评估,以验证该方法的有效性。研究结果表明,人工生成雾天场景车道线图片以扩充数据集的方法能够在薄雾情况下将雾天车道线的识别率从74.65%提升至86.65%,在中度雾下从51.41%提升至81.53%,在浓雾下从11.09%提升至70.41%。

关键词: 车道线检测, 深度学习, 雾天图片生成, CULane数据集扩充, FoggyCULane, 空间卷积神经网络(SCNN)

Abstract: In order to improve the lane detection accuracy of the deep learning algorithm under foggy condition, one of the effective methods is expanding the foggy lane dataset. In this paper, it expands one of the most widely applied datasets, CULane, into FoggyCULane, by artificially generating foggy lane images based on sunny-scene lane images in CULane. This work generates lane images under three different concentrations of fog according to the single image depth prediction and atmospheric scattering model. The foggy lane images are then combined with the lane label of the original images under sunny condition, thus it artificially establishes a new lane dataset named FoggyCULane, containing 107 451 labeled foggy lane images, which is 1.8 times larger than the original CULane dataset. Subsequently, the original CULane dataset and the FoggyCULane dataset are used to train the spatial CNN(SCNN) lane detection neural network, and the training results are evaluated in 12 types of complex lane scenes, which contains 3 different concentrations of foggy scenes, to verify the effectiveness of this method. The results show that compared with the original CULane, FoggyCULane can significantly increase the detection accuracy of lanes in foggy scene, which the detection accuracy increasing from 74.65% to 86.65% under thin mist condition, from 51.41% to 81.53% under moderate fog condition, and from 11.09% to 70.41% under dense fog condition.

Key words: lane detection, deep learning, foggy image generation, expansion of CULane dataset, FoggyCULane, spatial convolutional neural network(SCNN)