Image-based Road Surface Condition Detection Using Transfer Learning

Accurate prediction of road surface conditions can help authorities manage vehicular transportation effectively in large cities by helping to reduce congestion and the risk of accidents due to adverse weather. Image-based classification, using Convolutional Neural Networks (CNN) in combination with Transfer Learning (TL), can provide a real-time, data-driven and cost-effective solution for classifying road surfaces under varying weather, traffic and image recording conditions. This paper proposes an image classification approach leveraging TL with the ResNet50 architecture, enabling the efficient utilisation of pre-existing knowledge within deep neural networks to facilitate rapid model adaptation without extensive dataset collection. This study focuses on the challenging tropical conditions of Singapore as a use case where the performance of the proposed approach is evaluated on two distinct video footage/image datasets, namely fixed expressway cameras and dashcams. The strategy of coupling pre-trained models with TL is shown to consistently converge to better results in a faster time compared to training from scratch or with fine-tuning. The results provide valuable insights for traffic management authorities in selecting scalable architectures and training strategies for big data curation from traffic cameras, considering computational constraints for real-world deployment.