Most drivers are well aware of the growing threat to their vehicles from potholes. These invasive creatures have multiplied rapidly on our roads, wearing out tyre tread, cracking axles, and in some cases causing serious accidents. The costs are enormous, but those who manage and maintain the roads have limited resources to deal with the problem, even when the accident and damage risk is on the rise.
Research in the International Journal of Information and Decision Sciences has demonstrated a highly accurate, camera-based system for detecting and pinpointing potholes on roads. The work offers a new approach to how roads might be monitored and maintained to support critical transport infrastructure. It addresses the longstanding problem of identifying road surface damage soon enough to prevent vehicle damage and accidents.
Potholes usually form when a small area of the road is damaged in some way and water ingress weakens the layers beneath the road surface. This is exacerbated by freezing temperatures as that water expands and causes even more damage. If such a pothole is repaired well and in a timely manner, then the problems can be reduced.
Unfortunately, many potholes reach a hazardous size before they are even reported and so smaller potholes form near the primary holes as hardcore and gravel spewed from the initial spread across the road surface and are ground into the surface by vehicles. Similar problems subsequent to eventual repair also arise if the maintenance crew is limited to repairing only the main area and not any adjacent damage and if they fail to remove debris from the patched area and its surroundings before leaving.
Potholes, unfortunately, are not simply a matter of inconvenience, they can lead to significant damage to vehicles as well as risking serious accidents as drivers attempt to avoid them or else temporarily lose control of their vehicle when snarled by a deep pothole. Current inspection regimes are limited, wholly manual, and commonly rely on reports of damage, often after a driver and their vehicle have come unstuck.
The new approach uses ordinary cameras rather than specialised sensors to scan the road surface. Artificial intelligence, AI in the form of a deep learning model and a bespoke convolutional neural network, can analyse and classify the images based on visual patterns associated with the presence of potholes in the photos on which the system is trained.
Tests on real-world photos of roads under varying conditions of repair and disrepair showed classification accuracy of almost 99% with a good balance between false positives and negatives. Missed potholes leave hazards unaddressed, while false alarms waste maintenance resources. The new model could lead to a much smoother ride.
Dhiman, A., Kumar, M., Yadav, A.K. and Yadav, D. (2025) ‘Pothole detection and localisation from images using deep learning‘, Int. J. Information and Decision Sciences, Vol. 17, No. 4, pp.357-370.
No comments:
Post a Comment