Determination of the Pavement Surface Degradation Index Using the Instance Segmentation Method

Abstract

A nation’s development depends in part on the quality of its road network. To this end, resources are being deployed to monitor the surface condition of our roads. Applying deep learning to road damage detection can significantly optimize roadway monitoring campaigns. According to the VIZIR method, roadway diagnosis cannot be performed without assessing the degradation index after monitoring. This work aims to develop an auscultation tool based on transfer learning, object tracking, and image processing to estimate the pavement deterioration index. To achieve this, the YOLOV11 instance segmentation and Roboflow instance segmentation 3.0 models were trained on five databases compiled from videos of degraded road surfaces taken on various roads in Benin. At the end of these various training sessions, the best model, named RIS3.5, obtained after training on the fifth database converted to grayscale and comprising 19 classes, had an accuracy of 95%, a mAP of 94.8%, and a recall of 90%. This model was then used to track objects in real time and enable the assessment of extent and SDI through a Python script every 5.5 m and then every 50 m, 200 m, and 500 m.