AI Road Surface & Pothole Detection

Roads degrade continuously, but inspection budgets are fixed. Every road authority faces the same squeeze: a network that needs constant attention, a maintenance crew that can only be in one place at a time, and a public that notices every pothole long before the next survey is scheduled. AI road surface and pothole detection closes that gap. It turns ordinary road footage — from a maintenance van, a service vehicle, even a dashcam — into a live, prioritised map of potholes, cracks and obstacles, so teams fix the right defects first instead of the ones that happen to be reported loudest.

This page is an honest capability deep-dive, not a client success story. We explain the problem we solve, exactly how we would build the system, the technology behind it, who it is for and the return it can deliver — and we are deliberately careful with the numbers. The figures further down are sector benchmarks from peer-reviewed studies, not results from a specific Crux Digits route. For your network, we report your real numbers from a pilot on your own footage.

The problem: inspection that is always slightly out of date

Periodic manual surveys are expensive, slow and out of date the moment they are filed. A specialist survey vehicle is costly to commission and covers only a fraction of the network per pass, so most roads are assessed on a cycle measured in months or years. In between, defects grow, small cracks open into potholes, and the maintenance team works from a snapshot that no longer matches the road.

The naive fix — point a camera at the road and run a detector on each frame — fails in a predictable way. Single-frame detectors mistake shadows, wet patches, tar repairs, manhole covers and drain grates for damage, drowning the team in false positives until they stop trusting the system. A road authority does not need more alerts. It needs a continuous, trustworthy, prioritised view of the whole network, captured from vehicles it already runs.

How the system works

We build the pipeline in four stages, each designed to remove a specific failure mode rather than to add another model for its own sake. The result is a system that is reliable enough to act on, not just a demo that looks good on a clean test set.

• Capture — Multi-view video is recorded with GPS sync from vehicle-mounted or roadside cameras, then stabilised, deblurred and colour-normalised so that motion, vibration and changing light do not corrupt the input before detection even begins.
• Detect — Two complementary models run together. Semantic segmentation finds cracks and surface wear at the pixel level, while an object detector — typically YOLOv8 — finds potholes and debris, trained on dust, glare and motion so it holds up in real road conditions, not lab conditions.
• Filter — This is where most systems fall down and ours earns its keep. Consecutive-frame comparison and optical flow suppress shadows and wet patches: a real defect is physically present across frames as the vehicle moves, while a shadow is not. Temporal filtering is what turns a noisy detector into a dependable one.
• Map & rank — Every confirmed defect is pinned to GPS coordinates and graded by severity. Dashboards then show heatmaps and maintenance priorities, so the highest-risk repairs surface first instead of arriving as a flat, undifferentiated list.

That last stage is the difference between detection and decision support. Finding a pothole is the easy part; telling a maintenance planner which pothole to fix on which road this week — and proving why — is the part that actually changes how a road authority operates. This is the same applied, production-first approach we bring to all of our computer vision work.

The technology and approach

The detection backbone combines real-time object detection (the YOLO family) with semantic segmentation for fine cracking, because potholes and hairline cracks are genuinely different problems that reward different models. On top of that sits the temporal layer — frame-to-frame tracking and optical flow — which is the single biggest driver of real-world reliability and the hardest part to get right.

Around the models we build the unglamorous engineering that decides whether a system survives contact with reality: GPS and timestamp synchronisation, video stabilisation and normalisation, a severity-grading rule layer, and a mapping and dashboard front end for planners. Models can run at the edge inside the vehicle for live capture, or in batch on uploaded footage when you simply want a periodic network sweep. That data and pipeline work is its own discipline — the kind of robust capture, storage and processing our data engineering team specialises in — and it is what keeps the system trustworthy at scale rather than impressive only on a single clip.

Who it is for and the ROI

This system is built for road authorities and the organisations that serve them — municipal and provincial highway departments, national road agencies, and the contractors and asset managers who maintain networks on their behalf. It is a strong fit for any public-sector body accountable for road condition and safety, and for fleet operators who already drive the network every day and could be capturing condition data as a by-product of routes they run anyway.

The return comes from three places, and they compound. First, cost: condition data is gathered from vehicles already on the road, so there is no specialist survey fleet to commission. Second, prioritisation: a maintenance budget spent on the highest-severity, highest-risk defects first goes further than the same budget spent reacting to complaints. Third, safety and liability: a continuously updated, GPS-mapped, time-stamped record of defects and their severity is exactly the evidence a road authority needs to show diligence and defend decisions. Catching defects earlier, when they are cheap cracks rather than expensive potholes, is where preventive maintenance quietly pays for itself.

An honest word on the numbers

We will not put a fabricated client metric on this page. The benchmark figures below come from peer-reviewed road-defect studies published between 2022 and 2025: real-time YOLO models report pothole-detection mAP of roughly 81–86%, tuned and augmented models reach up to ~95%, and systems run at around 20 FPS with detection out to roughly 100 metres. These are sector benchmarks for the technology, not Crux Digits' own route results. Performance on your network depends on your cameras, your road types, your weather and your traffic — which is exactly why we measure it directly. For machine-learning systems like this, honest evaluation on your own data is not optional; it is the whole point, and it is central to how our machine learning practice works.

How we would run a pilot

We start small and prove it before anyone scales it. A focused route trial keeps the risk and the cost low while producing the only numbers that actually matter — yours.

• Scope a representative route — a stretch of the network that reflects your real conditions: road types, surfaces, lighting and weather.
• Capture on your own vehicles — mount cameras with GPS sync on vehicles you already run, so the trial reflects how the system would actually operate.
• Tune and validate — adapt the detectors to your footage and measure mAP and false-positive rates against ground truth, so you see real accuracy, not a generic benchmark.
• Review the prioritised map together — walk through the GPS heatmap and severity ranking with your maintenance planners to confirm it fits how they actually work.
• Decide on rollout with evidence in hand — scale across the network only once the pilot has proven the numbers and the workflow on your own roads.

If continuous, prioritised visibility of your road network sounds like the problem you are trying to solve, that is exactly the kind of focused pilot we scope. Tell us about your network and we will report real numbers, not promises — book a free consultation or see transparent scoping on our pricing page. For related computer-vision systems, see our work on concrete crack detection and low-emission-zone licence plate recognition.