AUTOMATIC ROAD ROUGHNESS DETECTION AND RANKING رسالة دكتوراة

Abstract

Road roughness is a significant challenge for government institutions, often requiring com plex processes, expensive tools, and significant manpower for evaluation. Traditional methods either classify roads as rough or not without quantifying the degree of roughness, or rely on costly sensor-based devices.Therefore, to solve these limitations, we created a model that catego rizes roads into six classes based on the International Roughness Index (IRI) and also addresses traditional issues related to measuring roughness, such as sensor costs.So, we have relied on drones to take the data for classification, because drones provide a clear, high-resolution view of the road surface, free of terrain or obstruction effects. In contrast, accelerometers and cameras installed on the vehicle measure road roughness based on the vehicle’s response, which can be inconsistent due to road conditions and external factors. Then, we proposed a novelty road roughness detection model utilizing drone videos and forest of hybrid computer vision and deep learning models to classify and determine the degree of road roughness. The proposed model uses drone-captured images of roads, which are processed and analyzed by a model of five neural networks, each specializing in different aspects of image analysis. The first network orders images from simplest to most complex based on GLSZM properties. The second network applies K-means clustering, using GLSZM features to optimize K values. The third network utilizes GLCM features for K-means clustering. The fourth network converts GLSZM into graph representations for classification. The fifth network combines K-means and LBP techniques to xiii enhance pattern visibility. Results from these networks are combined using a multi-class SVM to determine the final classification. We created a dataset consisting of 15,326 images taken from 1,500 videos recorded using a drone in 4K resolution, with a total runtime of 5,112 minutes, covering 81 km of road length. The dataset is divided into six classes, from degree 1 to 6 of road roughness based on IRI, and used this dataset to test our model. Our model of forest of hybrid computer vision and deep learning models achieved an accuracy of 95.665%, demonstrating its effectiveness in providing precise road roughness evaluations from drone imagery.