A Convolutional Neural Network Approach for Road Anomalies Detection in Bangladesh with Image Thresholding

Machine Learning is becoming an important tool for road condition detection. It is used in Road Information System (RiS), Road Weather Information System (RWiS), Intelligent Transportation System (ITS). All these systems are based on image classification and detection. Road images vary country wise. This paper aims to propose a road condition detection system for less developed countries such as Bangladesh. The research work presents a novel approach of using the Fully Convolutional Neural Network (FCNN) algorithm to detect the types of roads from street images of Bangladesh, detect the anomalies of the roads and classify the roads accordingly with black and white image thresholding. Based on the dataset of the roads of Bangladesh, a model has been trained to detect five classes of roads of Bangladesh. The model has been tested and validated by roads of other countries and showed an average accuracy of 87% for all classes. The goal of the research is to detect the condition of streets or roads of Bangladesh, based on structural and surface condition, material characteristics and classify them accordingly, from perfect roads to severely damaged road.