Road Defect Detection Using Deep Learning

Deteriorating roads pose safety risks to road users and can cause costly damage to the vehicles. The severity of hazards caused by road defects can range from minor to severe. These hazards can be minimised by the timely detection of road defects. Technological advancements have led to traditional inspection methods such as visual inspection being replaced by more advanced methods such as deep learning techniques for autonomous road defect detection. However, one of the major challenges faced by deep learning techniques is the requirement of significant amounts of training data. The acquisition of large amounts of data is rather costly due to equipment, vehicle fuel and data storage expenses. Additionally, integrating deep learning models for road defect detection with existing international codes and standards remains a challenge. This short paper presents a quicker and more efficient data acquisition method for acquiring data to train a deep learning road defect detection model using transfer learning. The model was also designed to allow for easy integration with the UK highway inspection manual. The model demonstrated good performance, achieving precision, recall and mAP values of 89.5%, 81.6% and 84.6%, respectively.