Pothole Detection and Classification in Vehicular Networks using ResNet-18

The impairment of road performance and structure due to insufficient serviceability for traffic is referred to as road damage. To mitigate this issue, several local governing bodies and transportation departments strive to incorporate automated methods for evaluating the extent of such damages on roads. Nevertheless, they frequently encounter limitations in terms of resources like cutting-edge technology, expertise, or financial capabilities required for collecting data and conducting an analysis on these road impairments. The GRDDC (Global Road Damage Detection Challenge) dataset and the ResNet-18 architecture are used to offer a Pothole detection and classification framework in this study. The goal is to group snapshots of road surface degradation into many categories, including cracks, potholes, and patches. It is used as the ResNet-18 design is a variation of the Residual Network, can accommodate deep layer structures employing skip connections. By initialising the ResNet-18 model using pre-trained weights from ImageNet, transfer learning is implemented, improving convergence speed and overall performance by utilising learnt features. Based on assessment results, hyperparameter adjustment and fine-tuning were carried out to enhance the model’s performance. To do this, experiments with architectural changes and hyperparameter adjustments to learning rate, batch size, and regularisation methods were conducted. To verify the robustness of the model, cross-validation methods were used.