Road hazard monitoring and severity assessment through computer vision techniques

Road hazards like potholes constitute a serious risk for both vehicles and transport infrastructure, which needs constant monitoring and maintenance. Automatic pothole detection from images is a challenging task due to the different environment conditions and variability like lighting, traffic, road type and materials, camera properties and so on. In this thesis, we propose a method to automatically detect dangerous potholes and to assign a severity score that reflects their potential impact on the vehicle. The approach uses images collected from an on-board vehicle camera to identify hazardous potholes and determine their locations, with a focus on vehicle and tire safety. We addressed this issue by collecting a bespoke dataset with more than 17000 road images with over 24000 instances of potholes located across the Italian regions of Lazio, Marche and Umbria. Many instances of the dataset were collected with a tailor-made smartphone application providing both images and vehicle telemetry features related to the collision with the hazard. To address the task we implemented deep learning supervised algorithms addressing both object detection to locate the hazard and a regression technique to predict the severity of the pothole for the vehicle and driver safety. The scope of this work is to provide an efficient solution that could also run on edge devices for detecting potholes in real time thus creating a pro-active monitoring system to address the health condition of the road infrastructure.