Abstract

The International Road Assessment Program (iRAP) relies heavily on manual video interpretation and manual coding of road attributes to evaluate the level of protection that road infrastructure provides to its users, which limits scalability, consistency, and efficiency. This study presents an automated framework that integrates deep-learning–based computer vision with the iRAP methodology to streamline the collection and processing of key road safety attributes. The proposed approach develops advanced models that enhance accuracy and enable efficient large-scale assessment of speed-limit signage, road-geometry characteristics, and roadside land-use transitions The proposed framework automates the detection and classification of critical road features by employing cutting-edge deep learning models including YOLO (You Only Look Once), Mask R-CNN (Mask Region-based Convolutional Neural Network) for object detection and instance segmentation, and DeepLabv3+ for semantic segmentation. To achieve high precision and generalization across various environmental conditions, these models are trained on diverse datasets, including street-view images and high-resolution aerial imagery. Key performance metrics such as precision, recall, F1 score, and mean Average Precision (mAP) are used to evaluate model efficacy, with initial results demonstrating promising improvements in detection accuracy and scalability. The findings of this research contribute to developing a cost-effective, automated solution for road safety assessments. This system significantly reduces human intervention, enhances the reliability of iRAP ratings, and supports informed decision-making to prioritize interventions for safer transportation networks. The framework aligns with global road safety goals by enabling timely and accurate assessments that reduce road traffic injuries and fatalities. Future work aims to further refine the model's accuracy and explore its application in diverse geographical contexts.