Unpaved Road Damage Detection via Semantic Segmentation: Enabling Smarter ADAS

The visual assessment of damages on unpaved roads, which lack the structured features of paved roads, is one of the primary issues facing the Advanced Driver Assistance System (ADAS). Most existing solutions focus on well-maintained roads with high-quality data and neglect the vehicles operating on unpaved or unstructured roads. This paper proposes a novel deep-learning approach for damage detection on unpaved roads using a UNet architecture with a ResNet-18 backbone. The encoder is ResNet-18, a lightweight network that uses residual connections to capture hierarchical information and alleviate the vanishing gradient issue effectively. Experiments are conducted on the Road Transverse Knowledge (RTK) dataset, which contains low-resolution images and challenging road conditions. The experiments are performed using three architectures—U-Net, LinkNet, and Feature Pyramid Network (FPN)—each tested with ResNet-18, InceptionV3, and VGG16 as the backbone. The results demonstrate that the UNet with ResNet 18 backbone achieves an IoU value of 73.92(%), and F1 Score value of 82.98(%). The U-Net with ResNet-18 backbone model outperforms other state-of-the-art (SOTA) models in segmenting damages on unpaved roads.