About the research
Safety is a principal concern for highway transportation, and slippery roads can pose high risks for vehicle crashes in snowy regions, which cover about 70% of road networks in the United States. Slippery road conditions can significantly increase the risk of vehicle crashes. Therefore, roadway agency staff find it critical to clear road surfaces in time to ensure traffic safety during ice and snow seasons. Moreover, the capability to estimate multi-lane roadway snow coverage is instrumental for snow plowing performance evaluation and resource planning for snowy regions during winter seasons.
The researchers developed and evaluated a sensing technology to evaluate multi-lane roadway snow coverage leveraging non-invasive dual-spectrum cameras, computer vision, and machine learning algorithms. The use of optical and infrared images for slippery roadway condition detection has the potential to operate in different illumination conditions.
The team deployed two dual-spectrum cameras, which can acquire both optical and infrared images of roadways. Computer vision algorithms were developed to perform image registration, segmentation, lane splitting, classification, and clustering.
Furthermore, to account for the relatively limited data volume, the researchers established a transfer learning framework, which greatly eliminated the need for training a large number of hyperparameters. The transfer learning algorithm achieved a precision of 88% using daytime optical images and an impressive precision of 94% when using nighttime thermal images, despite the constraints imposed by using a limited dataset.