A Comparative Analysis of Deep Learning Approaches for Road Anomaly Detection

Monitoring road pavement conditions is vital to prevent harm or accidents to vehicles and people, and it is a crucial governmental task. Some Artificial Intelligence (AI) based systems for automatically detecting and classifying road anomalies have been proposed in the literature. In particular, vision-based Deep Learning techniques process images to analyze the road pavement and detect different types of anomalies, thus offering a very flexible approach to road monitoring. This paper focuses on the comparison of two different vision-based techniques aimed at pothole detection: CNN and R-CNN. We have carried out our experiments by analyzing video streams acquired with a smartphone mounted on the windshield of a vehicle through a car mobile holder. Also, the Deep Learning classification models have been run on different computing infrastructures, such as a Microsoft Azure virtual machine (VM) and an edge device Raspberry Pi, to evaluate the usability in the Cloud and on-road infrastructures.