Road damage detection: how AI finds potholes and cracks from existing cameras

AI road damage detection uses computer vision to find potholes, cracks and other pavement defects in images and video from cameras that already drive past or watch the road: dashcams on municipal and service vehicles, CCTV, phones and drones. Each defect is classified, graded for severity and pinned to a map, so maintenance teams can see the whole network rather than only the streets an inspector reached that month. It doesn’t replace engineering surveys of ride quality and structure, but it keeps the record of surface damage current.

Why road damage needs finding quickly

Potholes start as cracks. Water gets into cracked pavement, freezes and thaws, and the weight of passing traffic breaks the surface apart. Left alone, a crack that was cheap to seal becomes a hole that damages vehicles. AAA found that in 2021, 1 in 10 US drivers had pothole damage serious enough to need a repair, at an average of almost $600 a repair, costing drivers $26.5 billion in total.

The volumes are large. In 2021 alone, Saudi Arabia’s municipalities repaired 15.9 million square metres of potholes and 4 million metres of broken pavement. Every one of those had to be found and recorded before it could be fixed.

What AI can detect on roads

One public benchmark, RDD2022, contains 47,420 road images from Japan, India, the Czech Republic, Norway, the United States and China, labelled with more than 55,000 instances of four damage types:

• Longitudinal cracks, running along the direction of travel
• Transverse cracks, running across it
• Alligator cracks, interconnected cracking that looks like a reptile’s skin and signals structural failure
• Potholes

Production systems go further, because the defects a city has to manage are not only in the carriageway. A 2023 benchmark built from 34,460 images collected through Saudi Arabia’s municipal ministry is labelled for excavation barriers, potholes and dilapidated sidewalks. CamCom’s deployments also cover damaged footpaths, sidewalk obstructions, open manholes, damaged or missing traffic signs, and construction and excavation waste.

How AI road damage and pothole detection works

1. Capture from vehicles and cameras already on the road. Dashcams on municipal and service vehicles, such as refuse trucks, cover the network on their normal routes. Fixed CCTV, inspectors’ phones, citizen reports and drones fill the gaps.
2. Detect and classify. Computer vision models find each defect and label it by type.
3. Grade severity. Size and type decide urgency: a deep pothole in a traffic lane is more urgent than hairline cracking on a side street.
4. Locate and de-duplicate. Each detection is geotagged, and repeat sightings of the same pothole from different vehicles are merged into one record.
5. Route to the right team. The defect goes to the department or contractor responsible for that road, with the image as evidence.
6. Confirm the repair. The next vehicle to pass shows whether it was fixed, and the history shows which roads keep failing.

How it fits with pavement condition surveys

Road agencies already measure pavement condition, but differently. In the United States, the Federal Highway Administration reports ride quality using the International Roughness Index (IRI): below 95 inches per mile is good, 95 to 170 is fair, and above 170 is poor, alongside measures of cracking, rutting and faulting. Those measurements come from specialised survey equipment and are taken periodically.

Camera-based AI answers a different question: where is visible surface damage right now, and what needs fixing first? The two work together. Surveys set long-term maintenance programmes; continuous detection catches the pothole that opened last week.

What affects accuracy

• Lighting and weather: low sun, shadows, wet roads and night-time video
• Look-alikes: shadows, patched repairs, road joints and stains that resemble cracks or potholes
• Camera position and speed: mounting height, angle and motion blur
• Local road types: asphalt, concrete and unpaved surfaces look different, and a model trained in one country needs examples from yours

When you evaluate a system, ask for accuracy by category from live deployments on roads like yours, and for how repeat detections are merged. A pothole reported 40 times is one work order, not 40.

Road damage detection at national scale: Saudi Arabia

Saudi Arabia’s Ministry of Municipalities and Housing, known until July 2024 as MoMRAH, uses CamCom’s AI platform to detect and classify public safety hazards, from road damage to visual pollution, in images from field surveys and citizens’ mobile reports. Each incident is classified, mapped to the relevant regulation and assigned to the responsible department, with alerts to officers and a national view in a central command centre.

Results CamCom has published from the deployment:

• 450,000 km of road documented, with an image every 50 metres
• 30 million+ incidents processed across 16 municipalities
• 43+ categories detected, up from 6 at launch
• Case processing cut from three weeks to under one week
• A 62% reduction in fatalities and casualties

Read the Saudi Arabia case study, and our explainer on visual pollution and how AI detects it.

For cities and road authorities, CamCom offers this as Vigil by CamCom: AI hazard detection for cities, using the cameras a city already has.

Common questions

How does AI detect potholes?
Computer vision models analyse images and video from dashcams, CCTV, phones or drones, find potholes and other defects, classify them by type and severity, and geotag each one so it can be mapped and sent to the team that repairs it.

What types of road damage can AI detect?
Longitudinal, transverse and alligator cracks and potholes are the standard categories in public datasets such as RDD2022. Production systems such as CamCom’s also detect damaged footpaths, sidewalk obstructions, open manholes, damaged or missing traffic signs, and construction and excavation waste.

What cameras are used for road damage detection?
Mostly cameras that are already on the road: dashcams on municipal and service vehicles, fixed CCTV, inspectors’ and citizens’ phones, and drones. Starting usually doesn’t need new hardware.

Does AI road damage detection replace pavement surveys?
No. Pavement surveys measure ride quality and structural condition periodically with specialised equipment. AI detection finds visible surface damage continuously, so the two are used together.

Related reading: Visual pollution: what it is and how AI detects it

Sources: AAA, Potholes pack a punch as drivers pay $26.5 billion in related vehicle repairs, 1 March 2022. Federal Highway Administration, Conditions and Performance report, chapter 6. Deeksha Arya et al., RDD2022: a multi-national image dataset for automatic road damage detection, 2022. Ministry of Municipalities and Housing, 2021 visual distortion figures (Arabic), 15 May 2022. Data in Brief, Visual pollution real images benchmark dataset on the public roads, 2023. CamCom, Saudi Arabia case study.