Real-Time Detection of Blowing Snow Events on Rural Mountainous Freeways Using Existing Webcam Infrastructure and Convolutional Neural Networks

The main objective of this study is to automatically detect real-time snow-related road-surface conditions utilizing existing webcams along interstate freeways. Blowing snow is considered one of the most critical road surface conditions, causing vertigo and adversely affecting vehicle performance. A comprehensive image reduction process was performed to extract two distinct reference datasets. The first dataset comprised two image categories: blowing snow and no blowing snow, while the second dataset consisted of five categories: blowing snow, dry, slushy, snow-patched, and snow-covered. Six pre-trained convolutional neural networks (CNN) were utilized for road-surface condition classification: AlexNet, SqueezeNet, ShuffleNet, ResNet18, GoogleNet, and ResNet50. In Dataset 1, it was concluded that AlexNet is a superior model with respect to training time and 97.56% overall detection accuracy. Regardless of differences in training time, ResNet50 achieved the highest overall accuracy of 97.88%, as well as the highest recall and F1-score. In Dataset 2, the ResNet18 model achieved an optimal overall detection accuracy of 96.10%, while the AlexNet model demonstrated the shortest training time and an overall detection accuracy of 95.88%. In addition, a comprehensive comparison was conducted between pre-trained CNNs and traditional machine learning models, with the former displaying significantly superior detection performance. Analysis of the confusion matrices revealed that AlexNet performed the best in detecting blowing snow events. The proposed models could automatically provide real-time accurate and consistent surface condition information.