Design of a deep learning based pothole detection system.

Abstract

With the number of potholes on the roads rising, as a result of various factors such as road saging due to delayed repairs and poor maintenance has resulted in negative impacts like traffic congestion, road flooding during the rainy season, accidents and damages to vehicles, which deeply affect and haunt road users in so many ways for example drivers, motor-riders, cyclists and pedestrians at large.Therefore a solution to combat down these negative effects is necessary by automating the detection process of these potholes to quickly identify their location minimizing the amount of time it takes for these potholes to be fixed from the period they are formed and first discovered.Hence a design of a deep learning based pothole detection system.The design analysis is on the use of machine learning and convolutional neural networks involving object detection and computer vision for the study of road pothole detection and location in order to create a data set of pothole images which will then involve training a machine learning model and also creating a database of all the potholes that have been detected using the model.Data collection was carried out using a cell phone mounted on a car windscreen to take pictures in form of videos of various roads inside a moving vehicle. The videos are then trimmed into corresponding images containing potholes and their locations. Pre-processing of data was also carried out in which images were transferred to a COCO data set format, after which the data set was splitted into testing, training and evaluation datasets. For training we used the Google Efficient Det model algorithm because of its high efficiency with small datasets.The images go through layers and at each layer, optimization is carried out by extraction of specific image features.