AI road inspection uses cameras mounted on ordinary vehicles, GNSS positioning and computer vision to find potholes, cracks and damaged road assets across a whole network, then scores every road segment in GIS so maintenance is planned from measured condition instead of complaints. Swedish Technology builds these pipelines for roads authorities, municipalities and owners of large private road networks in the UAE, and keeps the imagery inside the owner’s own environment.

Request a Road Inspection Proof of Concept · Talk to an Engineer

Can AI detect potholes automatically?

Yes. A deep learning model trained on local road imagery detects potholes reliably from a moving vehicle in daylight. It also detects longitudinal, transverse and alligator cracking, ravelling, patching and edge failure, though fine cracking needs higher resolution and slower capture. Each detection has a class, a confidence score, a photo and a position on the road centreline. An engineer confirms it before it becomes a work order.

What the system detects

Category Classes
Pavement defects potholes, longitudinal and transverse cracks, alligator cracking, ravelling, patching, edge failure
Road assets (same survey pass) traffic signs, road markings, lighting columns, guardrails, drainage covers
Condition defect density and severity per segment, combined into a documented condition index
Change deterioration trend against previous surveys

How AI road inspection works

  1. Capture. A camera and GNSS receiver on an operations or inspection vehicle record geolocated frames while it drives its normal routes. Higher-specification setups add more cameras, an IMU and LiDAR.
  2. Quality filtering. Frames affected by blur, glare, rain or obstruction are removed.
  3. Detection. The model classifies and locates defects and road assets in each frame.
  4. Positioning. Detections are projected from the image to ground coordinates and snapped to the road centreline through linear referencing.
  5. Scoring. Duplicates across frames and passes are merged. Each segment gets a defect density, a severity profile and a condition score with a visible formula.
  6. Review. Engineers check flagged segments and a QA sample in a map application, then confirm or correct.
  7. Action. Confirmed defects create work orders in the maintenance system (for example IBM Maximo or SAP PM) with photo, class and location. Completed repairs come back and close the record.

Three ways to inspect a road network

Criteria Visual inspection Hired survey vehicle Owned AI camera pipeline
Frequency whatever the team can drive usually annual as often as your vehicles drive
Consistency varies by inspector consistent per survey consistent, model-versioned
Output notes, photos, spreadsheets report or proprietary dataset features in your own GIS
Traceability low scores often not traceable to images every score links to images, model version and reviewer
Updates between surveys complaints only none continuous
Upfront effort low procurement per survey pilot, model training, integration
Best fit small networks one-off baseline surveys networks that need continuous condition data

Accuracy methodology

A single accuracy number for road AI is not meaningful. During the pilot we:

  1. Survey a representative 20 to 50 km sample across road types and conditions.
  2. Label defects on locally captured frames with your engineers’ definitions.
  3. Measure precision and recall per defect class on a held-out section the model has not seen.
  4. Compare segment scores with your engineers’ own ranking of the same sections, and calibrate until the ranking holds.
  5. Report which classes are ready for automation and which stay manual.

Architecture

Layer Contents
Capture vehicle cameras, GNSS, optional IMU and LiDAR; recording app storing survey ID, speed and lane
Ingestion upload, quality filter, trajectory smoothing, face and number-plate blurring
Detection deep learning models on GPU servers (ArcGIS deep learning tools or a PyTorch pipeline), model registry
Scoring segment aggregation, configurable condition index, trend analysis
Review map-based review app with audit fields on every detection
Integration maintenance system work orders, dashboards, Arabic and English report packs

UAE considerations

  • Imagery stays local. Survey imagery covering an emirate is sensitive. Storage, inference and review run on-premises or in a UAE-region private cloud.
  • Surface behaviour. High surface temperatures and heavy vehicle loads produce rutting and bleeding patterns that models trained abroad miss.
  • Capture timing. Glare, blown sand and night driving reduce reliability. Survey windows are planned around them.
  • Privacy. Faces and number plates can be blurred at ingestion, and access to raw frames is limited to named reviewers.

Limitations

  • Cameras see the surface only. Structural capacity still needs deflection testing, coring or ground-penetrating radar.
  • An image-based condition index is comparable to, but not the same as, a formally surveyed ASTM D6433 PCI.
  • A poor road centreline breaks segment assignment, so it is fixed before detection work starts.
  • The output is prioritised evidence. Treatment decisions stay with your engineers.

Implementation steps

  1. Scoping (one to two weeks): network extent, segmentation, defect classes that drive treatment decisions.
  2. Data review: centreline quality, existing condition records, complaint data.
  3. Capture pilot (two to four weeks): one or two vehicles on a representative sample.
  4. Model development and per-class measurement.
  5. Condition index calibration against your engineers’ assessments.
  6. Review app and work-order integration.
  7. Network rollout with a survey frequency per road class.
  8. Handover: models, training data, configuration, code, runbook and retraining plan.

Related

Smart inspection AI · Road inspection and pavement condition in GIS (technical guide) · Road asset management with GIS · AI for roads and infrastructure inspection: project scope · Smart roads and transportation · Roads and transport

Questions buyers ask

How does AI road inspection work? Vehicle cameras record the road with GNSS positions, a computer vision model detects defects in the frames, GIS places each defect on the road segment, and engineers confirm findings before they become work orders.

Can we use our existing vehicles? Yes. A camera and GNSS receiver can be mounted on vehicles that already drive the network, which is what makes frequent surveys affordable.

How often can the network be surveyed? As often as the vehicles drive it. Most authorities choose a frequency per road class, for example more often on arterial roads.

Does it also inventory road signs and markings? Yes, from the same imagery, so a separate asset survey is not needed.

Does the imagery leave our network? No. Processing and storage run on-premises or in a UAE-region private cloud under your retention rules.

Can it integrate with our maintenance system? Yes. Confirmed defects create work orders with photo, class, severity and location. We build the integration for your system as part of the project.

What about public pothole complaints? Complaints are matched to the segment and its measured condition, so repeated reports on one street show up as a structural problem rather than as separate repairs.

Who owns the model? You do. Models, training data and code are handed over.

Request a road inspection proof of concept

Book a site survey or demo, or contact our engineers.

Solution scenarios

How this typically works for a specific kind of organisation. These are scenarios, not client case studies. See all solution scenarios.