ABSTRACT

Road Weather Information Systems (RWIS) provide real-time weather information at point locations and are often used to produce road weather forecasts and provide input for pavement forecast models. Compared to the prevalant street cameras, however, RWIS are sometimes limited in availability. Thus, extraction of road conditions data by computer vision can provide a complementary observational data source if it can be done quickly and on large scales. In this paper, we leverage state-of-the-art convolutional neural networks (CNN) in labeling images taken by street and highway cameras located across North America. The final training set included 47,000 images labeled with five classes: dry, wet, snow/ice, poor, and offline. The experiments tested different configurations of six CNNs. The EfficientNet-B4 framework was found to be most suitable to this problem, achieving validation accuracy of 90.6%, although EfficientNet-B0 achieved an accuracy of 90.3% with half the execution time. The classified images were then used to construct a map showing real-time road conditions at various camera locations. The proposed approach is presented in three parts: i) application pipeline, ii) description of the deep learning frameworks, iii) the dataset labeling process and the classification metrics.

摘要

道路天气信息系统 (RWIS) 提供点位置的实时天气信息，通常用于生成道路天气预报并为路面预测模型提供输入。然而，与流行的街头摄像机相比，RWIS 的可用性有时是有限的。因此，如果可以快速、大规模地提取道路状况数据，那么通过计算机视觉提取道路状况数据可以提供补充的观测数据源。在本文中，我们利用最先进的卷积神经网络 (CNN) 来标记北美街道和高速公路摄像机拍摄的图像。最终的训练集包括 47,000 张标记为五个类别的图像：干、湿、雪/冰、差和离线。实验测试了六个 CNN 的不同配置。EfficientNet-B4 框架被发现最适合这个问题，实现了 90.6% 的验证准确率，尽管 EfficientNet-B0 以一半的执行时间实现了 90.3% 的准确率。然后使用分类图像构建一张地图，显示各个摄像头位置的实时路况。所提出的方法分为三个部分：i) 应用程序管道，ii) 深度学习框架的描述，iii) 数据集标记过程和分类指标。