Opsio - Cloud and AI Solutions
Rail inspection vehicle at dusk on ballasted track, its under-car lamps lighting the rail heads, clip fastenings and concrete sleepers ahead

AI Visual Inspection for Infrastructure

Find rail head checks, concrete cracks and spalling, and pipeline corrosion in the images your inspection runs already capture, so your engineers review the worst defects first.

Book a walkthroughSee use cases

What is AI visual inspection for infrastructure?

AI visual inspection for infrastructure is the use of trained vision models to find, locate and grade defects in images of rail track, structures and pipelines. Opsio builds it for railway track inspection, bridge inspection and pipeline surveys, covering three defect families: rail and track defects, concrete and structural defects, and corrosion and coating defects.

Key facts

  • Rail imaging runs from vehicle-mounted line-scan cameras, with every image tied to track chainage.
  • Bridge and tunnel imagery comes from drones or vehicles and is mapped onto a photogrammetric 3D model.
  • Onboard edge inference runs under 50 ms per frame on NVIDIA Jetson or Intel OpenVINO hardware.
  • Each finding carries GNSS position, chainage or grid reference, severity grade, image hash and model version.
  • Findings export to enterprise asset management (EAM) and GIS systems as work requests and map layers.
  • Raw imagery stays on your premises or in a chosen EU cloud region.

Main defect families

Rail and track defects
Rolling contact fatigue cracks, spalling and corrugation on the rail head, plus fastening, sleeper and ballast faults.
Concrete and structural defects
Cracks, spalling, exposed rebar, efflorescence, bearing displacement and lining defects in bridges and tunnels.
Corrosion and coating defects
Metal loss and coating breakdown on steel members, pipelines and storage tanks, plus dents and third-party damage.

Last reviewed by Praveena Shenoy, Head of AI & Data, Opsio, a Swedish cloud and AI services company. Part of Opsio's AI Visual Inspection service.

Why is infrastructure inspection hard to scale?

Your networks are long, and your qualified engineers are few. One night's run on a main line can produce hundreds of thousands of rail images, and a drone survey of one viaduct several thousand photos, most of them reviewed by senior engineers.

The assets do not behave like parts on a production line. Lighting changes with weather, season and time of night, and surfaces carry rust, moss, graffiti and old repairs that look like defects but are not. Each rail profile, sleeper type, bridge material and coating adds variation.

Manual review breaks down on volume and consistency. After hours of near-identical frames, reviewers miss small squats or grade the same crack differently from one campaign to the next. Repeat surveys only help if you can compare like with like at the same location.

Escapes are expensive and public. A missed head check can grow into a broken rail, a missed spall can drop concrete onto a road, and a missed coating failure can end in a leak, followed by speed restrictions, closures and regulator scrutiny.

Railway track, bridge inspection and pipeline use cases

Three asset groups, each with its own campaign, imaging method and owner regime. In every case the system finds, locates and grades; your engineer confirms.

Railway track inspection

Close-up of a worn rail head with cracking and a piece of steel broken out at the gauge corner, next to a loose rail fastening clip lying on a concrete sleeper
Rail head with gauge-corner cracking and spalling, and a detached fastening clip on the sleeper

Situation

Your track inspection vehicle or a fitted service train runs the line, often at night, with cameras under the car. A typical run covers tens to hundreds of kilometres of continuous images of both rails, fastenings, sleepers and ballast.

Today, analysts scroll that footage or clear alerts from fixed rules that fire on every shadow and rust stain. Ultrasonic runs find internal flaws; the camera run finds surface and fastening faults, and that is where review time piles up.

Defects

  • Head checksfine rolling contact fatigue cracks on the gauge cornerHigh wheel-rail contact stress in curves repeatedly deforms the surface, and the cracks can branch into the rail if not ground out.
  • Squatsa dark, widened spot on the running band with V-shaped cracksThese rolling contact fatigue defects start at a surface irregularity and grow under traffic, sometimes leading to rail breaks.
  • Surface spallingSmall pieces of rail head metal break out, usually where head check or squat cracks join below the surface.
  • Rail corrugationA periodic wave of bright peaks and dark troughs along the running band, caused by wheel-rail dynamics, traction and braking.
  • Missing or broken fasteners and clipsClips, bolts and insulators go missing or crack from fatigue, vibration or poor installation, letting the rail move.
  • Sleeper cracksConcrete sleepers crack at the rail seat or centre from impact loads and poor support; timber sleepers split and decay.
  • Ballast foulingFine material from abrasion, subgrade pumping or spilled loads fills the voids between stones and shows as muddy, discoloured ballast.

Inspection setup

Two line-scan cameras per rail (one top-down over the head, one angled at the fastenings) sit in sealed housings under the vehicle, with C-mount lenses and LED line lights. Structured light (projected laser lines) adds a 3D profile for corrugation, missing clips and ballast level. A wheel encoder triggers each line, so resolution stays at typically 0.5 to 1 mm per pixel at any speed.

Every image is referenced to chainage (distance along the line from a fixed datum) and GNSS, and each location is flagged for review, logged for trend, or passed. The same encoder-synchronised method drives continuous strip and web inspection.

Outcome

Your analysts review a short, ranked list of flagged locations instead of the whole run. Each flag carries chainage, rail side, an image crop, a proposed class and a severity grade from your scheme. Repeat runs line up by chainage, so you see whether a squat is growing. Edge vision processing keeps raw data off the mobile link.

Bridge and tunnel inspection

Round concrete bridge pier with map cracking, a spalled area exposing corroded reinforcement bars, white efflorescence streaks and rust staining running down the face
Concrete bridge pier with map cracking, spalling, exposed rebar and efflorescence

Situation

Your bridges are inspected on a fixed cycle set by the owner, typically general inspections every one to two years and principal ones every six. Inspectors use platforms, rope access or road closures, and more now fly drones for close-range photos of piers, soffits and bearings.

Tunnels are inspected from a road or rail vehicle at night. Capture is getting easier; review is not, as one structure can produce thousands of images to sort and compare with the last report.

Defects

  • Concrete cracksCracks form from shrinkage, thermal movement, overload or rebar corrosion; width estimation in millimetres matters because owners grade severity by width and pattern.
  • SpallingSpalling is concrete breaking away from the surface, usually because corroding reinforcement expands, or from freeze-thaw and impact.
  • Exposed rebarReinforcement becomes visible where cover is lost through spalling or was too thin, and it corrodes faster once exposed.
  • Efflorescence and water ingressWhite salt deposits and damp staining show water moving through cracks and joints, carrying dissolved minerals out of the concrete.
  • Corrosion of steel membersGirders, bracing and connections lose section where coatings fail and water and de-icing salts collect.
  • Bearing displacementBearings shift or run out of travel from thermal movement, settlement or seized sliding surfaces, visible as offset plates.
  • Tunnel lining cracks and segment joint defectsLinings crack under ground load and ageing; segment joints open, step or leak when gaskets or bolts fail.

Inspection setup

For drone bridge inspection, a 45 to 61 MP RGB camera on a gimbal flies planned paths with typically 70 to 80 percent overlap, giving a ground sampling distance (GSD, the size of one pixel on the surface) of typically 0.2 to 0.5 mm. Photogrammetry builds a 3D model, and each crack, spall and stain is traced with semantic segmentation and projected onto it. Widths below one pixel are estimated from intensity profiles and flagged as lower confidence.

Tunnels use vehicle-mounted line-scan or area cameras with LED floodlights, encoder-linked to chainage and ring number. Processing is offline; each finding is flagged with a proposed condition grade for the engineer to confirm.

Outcome

Your engineers start the report with every visible defect already located on the model and sorted by proposed severity. Comparison between campaigns highlights new cracks, growing spalls and fresh efflorescence at the same location. Your engineer still signs the inspection. The same drone workflow serves wind blades and power lines.

Pipeline and storage inspection

Above-ground coated steel pipeline on a gravel right-of-way with a patch of blistered, disbonded grey coating exposing corroded steel beneath
Coated steel pipeline with coating disbondment and external corrosion on a gravel right-of-way

Situation

Your pipeline right-of-way is patrolled by air, drone or on foot, typically every few weeks to every few months. Above-ground pipe, tank shells and roofs at storage sites follow their own cycle.

Internal condition comes from inline inspection (ILI) runs with smart pigs carrying magnetic flux leakage (MFL) and ultrasonic sensors, and from crawler cameras on smaller lines and tanks. Each source arrives separately, and the integrity engineer correlates them by hand.

Defects

  • External corrosionMetal loss on the outer wall where coating has failed and moisture, soil chemistry or stray current attack the steel.
  • Coating disbondmentThe coating loses adhesion and lifts or blisters, from poor surface preparation, cathodic disbonding or soil stress, and shields steel from cathodic protection.
  • DentsLocal deformations of the pipe wall from rock impingement, settlement or impact that concentrate stress.
  • Third-party damageScrapes, gouges and dents from excavators and other unauthorised work near the line, the most common cause of serious incidents.
  • Vegetation and ground movement near the right-of-wayTree growth, erosion, landslips and subsidence block access, hide the line or strain the pipe.
  • Internal weld and corrosion indications from ILI imageryGirth weld anomalies and internal pitting show in crawler video and ILI-derived images, where flow, water dropout and deposits drive corrosion.

Inspection setup

Right-of-way patrols use drone RGB at typically 2 to 5 cm GSD, with optional thermal, plus satellite imagery for change detection between passes. Close-range photos of above-ground pipe and tank shells use raking light to show blisters and lifted coating. Crawler and pig footage is linked to odometer distance and girth weld number. Asset monitoring with computer vision also covers fixed cameras at valve stations.

Processing is offline. The system flags findings for engineer review, logs encroachment for the patrol team, and aligns visual findings with MFL and ultrasonic features. It never replaces the ILI data.

Outcome

Your integrity engineers see visual evidence and ILI features side by side for the same location. New excavation, vegetation and ground movement show up as change between passes, with a map position for the patrol team. Coating defects on above-ground pipe and tanks are graded and tracked over time, supporting your recoating plan without changing who decides.

How AI inspection works on an inspection run

Your asset is not a production line, so the walk-through follows an inspection run rather than a part on a conveyor. The steps are the same for a track vehicle, a drone mission or a pig run.

Machine-vision camera between two angled LED light bars mounted under a rail inspection vehicle, with a striped light pattern projected onto the rail and ballast below
Under-car camera and LED bars on a rail inspection vehicle projecting structured light onto the rail
  1. 1

    Trigger and asset presentation

    On rail, a wheel encoder triggers each image line so pixels map to fixed track distance. Drone flight plans trigger frames at set overlap; pig and crawler odometers stamp each frame with distance.

  2. 2

    Image acquisition

    GigE Vision line-scan or area cameras with C-mount lenses sit in IP67-rated housings with LED line lights or laser projectors. Drones carry a gimbal-stabilised RGB camera and, where needed, thermal.

  3. 3

    Edge inference and latency budget

    On vehicles, an NVIDIA Jetson or Intel OpenVINO computer runs inference under 50 ms per frame, keeping pace with acquisition. Drone and pig data run in batch after the mission, on premises or in an EU cloud region.

  4. 4

    Decision and actuation

    There is no reject gate: the decision is a graded flag queued for engineer review. Confirmed findings become EAM work requests via REST API, or alarms to vehicle systems over OPC UA or Modbus.

  5. 5

    Traceability and data retention

    Each raw image is hashed (SHA-256) at capture and kept in write-once storage with position, model version and reviewer decision, so the chain of custody for engineering decisions stays intact.

  6. 6

    Active-learning loop

    Every engineer confirmation or correction is stored with the image. Uncertain findings and new asset types go to a review queue; Opsio retrains on Amazon SageMaker and you approve each model version.

Which imaging finds which infrastructure defect?

Visual inspection covers surface-visible defects. Internal rail flaws, concrete delamination and internal wall loss still need ultrasonic, MFL, radar or hammer sounding.

Defect familyBest-suited imagingWhy
Rail defects: head checks, squats, surface spallingVehicle-mounted line-scan over the rail head, LED line light, encoder-triggeredConstant mm-per-pixel resolution shows fine crack patterns on the running band at line speed.
Rail corrugation and ballast levelLaser triangulation (structured light) 3D profile plus line-scanCorrugation and ballast shoulders are height features; 3D separates them from colour and rust.
Fasteners, clips and sleeper cracksAngled line-scan or area cameras, diffuse LED lightingAn oblique view shows clip presence and shape, and cracks beside the rail seat.
Concrete cracks, spalling, exposed rebar, efflorescenceDrone bridge inspection: high-resolution RGB plus photogrammetry, typically 0.2 to 0.5 mm GSDFine GSD resolves crack width; the 3D model locates and sizes each defect between campaigns.
Tunnel lining cracks and segment jointsVehicle-mounted line-scan or area cameras, LED floodlights, encoder-linkedAn unrolled image of the lining places each crack and joint step on the tunnel chainage.
Steel corrosion, bearing displacementClose-range drone or handheld RGB, raking light, fixed repeat viewpointsColour shows rust and coating loss; repeat viewpoints show bearing offset over time.
Pipeline corrosion and coating disbondment (external)Close-range RGB with raking light; thermal imaging for above-ground lines and tanksRaking light shows blisters and lifted edges; thermal can show moisture under coating.
Right-of-way encroachment, vegetation, ground movementDrone RGB (2 to 5 cm GSD) and satellite imagery, change detectionChange between dated passes is the signal.
Internal weld and corrosion indicationsCrawler and pig camera footage with MFL and ultrasonic ILI dataVideo shows the surface; MFL and ultrasonic measure wall loss.

Asset management, GIS and owner inspection regimes

Findings go into the systems your engineers already use, under your existing inspection rules.

Asset management (EAM/CMMS)
Reads asset IDs, track sections and previous defect records. Writes confirmed findings as defect records or work requests with image, grade and location, via REST API or file export.
GIS and linear referencing
Every finding carries coordinates and a linear reference (line and chainage, or pipeline station). Exports as GeoJSON, GeoPackage or shapefile, or writes to your GIS feature service.
Severity grading schemes
Configured to your scheme: track defect categories with action times, bridge element condition ratings or pipeline anomaly priorities. The system proposes; a qualified engineer sets the final grade. The system never closes out a defect or signs an inspection.
Owner inspection regimes
Designed to support statutory inspections such as the US National Bridge Inspection Standards, UK CS 450 for highway structures, FRA track safety standards and EN 13848 track geometry. Rail defects can follow the UIC 712 catalogue.
Chain of custody for images
Hash at capture, write-once storage, full metadata (time, position, camera, calibration, model version) and an audit log of every review decision.
Where the data lives
Raw imagery stays on the vehicle and your premises, or in an EU cloud region you choose. Access is role-based and logged.
Pipeline integrity data
Visual findings align with ILI feature lists by odometer distance and girth weld number, supporting integrity programmes that use API 1163 for ILI qualification.

From first test to your line in five steps

Infrastructure projects start with the images you already have: first tests run on past campaigns, before any hardware is fitted.

  1. 1

    Opsio runs tests in the AI Lab

    You share image sets from past campaigns: typically 2,000 to 10,000 images per asset type, with at least 50 to 200 confirmed defect examples per class plus defect-free images. Typical classes: head checks, squats, missing clips, concrete cracks, spalling and coating disbondment. Opsio trains a first model and measures detection rate and false-reject rate per class against your engineers' grades.

    You provide
    Image sets with past inspection reports; physical samples where useful, such as removed rail sections or broken clips.
    Opsio delivers
    Per-class results on your own images, with example findings.
  2. 2

    Present results and pricing

    Opsio walks your asset engineers through per-defect results, example finds, escapes and false rejects, and the review time per campaign they imply. You get the recommended camera, lighting and edge hardware for your vehicle or drone workflow, a check that processing keeps pace with your run speed, and a fixed quote for the system and managed service.

    You provide
    Asset, engineering and procurement reviewers.
    Opsio delivers
    Results report, hardware list and fixed quote.
  3. 3

    Decision and PO/MSA signing

    You decide. The purchase order and master services agreement set the scope (asset types, defect classes, networks), acceptance criteria per defect class, ownership of images and labels, the severity scheme to configure, data residency, chain-of-custody requirements for imagery and support terms.

    You provide
    Signed purchase order and a named asset owner contact.
    Opsio delivers
    Signed MSA, project plan and acceptance criteria.
  4. 4

    Installation in the Opsio AI Lab

    Opsio builds the production set in the AI Lab: camera housings, lighting, encoder interface and edge computer for a vehicle, or the processing pipeline for drone and pig data. It is tested against the acceptance criteria on held-out campaigns, and EAM and GIS exports are checked in your test environment.

    You provide
    Test access to EAM and GIS; vehicle mounting drawings.
    Opsio delivers
    Pre-commissioned hardware and pipeline, with test records.
  5. 5

    Installation at the customer location

    The kit is fitted to your inspection vehicle, or the pipeline connects to your drone workflow and asset systems. First campaigns run in shadow mode next to your current review, so engineers compare flags with their own findings. Then flagged review goes live, analysts are trained and the active-learning loop starts.

    You provide
    Vehicle or depot access, possessions or flight windows, analysts for shadow review.
    Opsio delivers
    Installation, shadow comparison, training and managed service.

Pricing for the feasibility study, production stations and the managed service is on the AI Visual Inspection service page.

AI visual inspection for Infrastructure: questions

What are the types of defects in rails?

Rail defects split into surface defects, internal defects and track component faults. Surface defects include head checks, squats, spalling, corrugation and wheel burns, and cameras find these well. Internal defects such as tache ovale and transverse fatigue cracks sit inside the rail and need ultrasonic testing. Component faults, such as missing clips, cracked sleepers and fouled ballast, are visual and suit automated imaging. The UIC 712 catalogue gives a common naming scheme.

What is it called when a train track splits?

A split or fractured rail is called a broken rail or rail break. It often starts as a fatigue crack, such as a squat or a transverse defect, that grows under traffic until the rail fractures. Visual inspection catches the surface-breaking cracks that can lead to a break, and a fully open break where the gap is visible. Internal cracks before a break need ultrasonic testing, so the two methods work together.

What causes concrete spalling?

Concrete spalling is mostly caused by corroding reinforcement, which expands and pushes the cover concrete off. Chlorides from de-icing salts or sea water and carbonation of the cover start that corrosion. Freeze-thaw cycles, impact, fire and poor cover depth also cause spalling. On images, the AI finds the spall outline, any exposed rebar and nearby rust staining, and measures the spalled area on the 3D model.

Is spalling concrete a structural problem?

Spalling concrete can be a structural problem, depending on where it is and how much steel is affected. Small surface spalls are often cosmetic, but spalls that expose corroding main reinforcement, sit at bearings or support zones, or keep growing need an engineer's assessment. The AI flags and sizes each spall and compares it with earlier campaigns; your structural engineer decides on the structural significance.

What is the difference between delamination and spalling in concrete?

Delamination is a hidden crack plane parallel to the surface, while spalling is concrete that has already broken off. Delamination usually comes first, as rebar corrosion splits the cover, and it is not visible in photos. It is found by hammer sounding, chain drag, infrared thermography or ground-penetrating radar. Visual AI detects the spalling, cracking and rust staining that often show where delamination is present.

When should you worry about concrete cracks?

Worry about concrete cracks when they are wide, growing, leaking or follow a structural pattern. Owners usually grade by width (often from about 0.3 mm upwards for reinforced concrete in exposed conditions), orientation, location and signs of water or rust. Diagonal cracks near supports and cracks along reinforcement lines matter more than fine shrinkage cracking. The AI measures width and length and tracks change between campaigns.

What are the concrete crack detection methods?

Concrete crack detection methods include visual inspection with crack gauges, photo and drone imaging, laser scanning, ultrasonic pulse velocity and acoustic emission. For network-scale bridge inspection, drone or vehicle photos with AI segmentation are the practical first pass. The model traces each crack as a line, estimates width from the pixel size, and projects it onto a 3D model. Your engineer then checks the flagged cracks on site or with a crack gauge.

What is a rail inspection vehicle?

A rail inspection vehicle is a train, road-rail vehicle or fitted service train that records track condition as it runs. Depending on the fit, it measures track geometry, runs ultrasonic rail flaw detection, and captures camera images of the rail, fastenings and sleepers. Opsio fits line-scan cameras, lighting and an edge computer to your existing vehicle, or processes images from the camera system it already carries.

What crack detection dataset do you need to train the model?

You need images from your own assets, not only a public crack detection dataset. Public datasets help pre-train a model, but your concrete finish, lighting, camera and grading scheme decide the result. Typically 50 to 200 confirmed examples per defect class, plus a larger set of defect-free images, are enough for a first feasibility model. Past inspection reports that link images to grades make the data much more useful.

What does rail grinding do?

Rail grinding removes a thin layer of the rail head to take out surface defects and restore the profile. It removes head checks and early squats before they grow, and it removes corrugation. AI inspection supports grinding planning by showing where head checks, squats and corrugation sit along the line and how they change between runs. Your track engineers use that map to plan grinding passes and check the result afterwards.

What is a pipeline inspection gauge (PIG)?

A pipeline inspection gauge, or pig, is a device pushed through a pipeline by the product flow to clean or inspect it. Smart pigs carry MFL, ultrasonic or geometry sensors that measure wall loss, cracks and dents. Some pigs and crawlers also carry cameras. Opsio analyses that camera footage and aligns visual indications with your ILI feature list by odometer distance; it does not replace the ILI measurement.

What are the bridge inspection requirements?

Bridge inspection requirements are set by the owner and the national regime. In the US, the National Bridge Inspection Standards set inspection types, intervals and inspector qualifications. In the UK, CS 450 sets general and principal inspections for highway structures. Most regimes require a qualified engineer to lead and sign the inspection. AI image analysis supports those inspections with located, graded findings; it does not change who is responsible.

How often does a bridge need to be inspected?

Most bridges need a routine inspection every one to two years and a detailed inspection every few years, depending on the regime. The US National Bridge Inspection Standards set a routine interval of up to 24 months for most bridges, shorter for structures in poor condition. UK highway structures typically get general inspections every two years and principal inspections every six. AI analysis makes each of those campaigns faster to review.

What kind of engineer inspects bridges?

Bridges are inspected by structural or civil engineers, or by trained bridge inspectors working under a qualified team leader. The exact qualifications depend on the country and the owner. AI visual inspection does not change that. It sorts and grades the image set so the qualified engineer spends their time on the defects that matter, then confirms or corrects every finding before it goes into the report.

Book a 20-minute walkthrough for infrastructure

Pick the defects that cost you most, and we will show you live how a vision model finds them on a line like yours.

  • Live on your defect type, no slides
  • A reply within one business day
  • No obligation
Your call is with
  • Vaishnavi ShreeDirector & MLOps Lead
  • Praveena ShenoyCountry Manager, India

We use your details to answer your request, as described in our Privacy Policy.

Other industries