Determination of the Degradation Index by Detection of Pavement Distress with Transfer Learning and Image Processing

Abstract

A road network is essential to a country’s transportation and socio-economic development. Its maintenance requires regular monitoring to guide maintenance decisions. Artificial intelligence now enables the automatic detection of damage, but monitoring a roadway does not end with detection. It also involves estimating the severity and extent of damage and determining the Is index. Therefore, this study allowed the development of a digital tool based on artificial intelligence and image processing for complete monitoring. This consists of detecting the roadway damages, then estimating their severity and extent to calculate the index Is . Four databases were designed from videos of damaged roads collected on different roads in Benin. These data were used in transfer learning to train the YOLOv9 and Roboflow 3.0 Object Detection models. A script was developed to estimate the extent of the degradations, setting a collection speed of 10 km/h, a picking height of 1.20 m and a viewing angle equal to 45˚, covering the entire width of the roadway. Another script determines the index Is by estimating the cracking and deformation indices, with possible corrections depending on the repairs present. The best model obtained, ROCNN4, results from training Roboflow 3.0 Object Detection with the fourth base. It detects 19 classes of degradations with a precision P of 90.8%, a mAP of 91.8% and a recall R of 89.5%. These results pave the way for better road maintenance planning by providing managers with a reliable and automated decision-support tool. They thus help optimize intervention costs and improve the durability of the road network.