Automatic Detection of Pavement Distresses Using Convolutional Neural Networks

Abstract

This paper presents a method for detecting pavement distresses using convolutional neural network architectures. The dataset used as training data was Road Damage Dataset 2022 (RDD2022) which has been adapted for the YOLO (You Only Look Once) architecture. In this study, bounding box detection and classification techniques were employed with several variants of YOLO and RT-DETR (Real-Time Detection Transformer) architecture. The global quality metric was mean Average Precision at 50% (mAP-50%), which varied between 64% and 70.4% depending on the model applied. Inference of the fine-tuned model was conducted on images excluded from the training dataset. All images subjected to detection were later geo-tagged and plotted on OpenStreetMap (OSM). Although this study showed promising results, it should be mentioned that more annotated data is required to achieve more precise results.