AI road inspection: where manual walking stops working
Manual road inspection rests on a human: a crew walks the lane and logs defects into a tablet, and the assessment depends on the shift, the lighting, and the experience of the specific inspector.
The same section gets different ratings on an overcast day and at noon, and the seasonality of the survey adds scatter — in spring cracks are open, by summer some hide under dust and traces of pothole patching. Data between surveys turns out to be incomparable, and repair cannot be planned on it.
Defect misses are most noticeable in ruts and at joints: that is exactly where the eye slides past, and a thin crack reads as part of the rut.
An individual joint passes inspection without remarks, and a season later turns into a pothole and emergency repair — the cost of work at that stage is already entirely different.
The second bottleneck is volume. The network expands faster than it can be walked: no matter how many crews you assign, only part gets covered in a season, and the rest is surveyed on a residual basis.
The problem is not the diligence of the crews, but the fact that the manual method does not scale to the entire network.
Here is how this gets closed: AI road inspection moves road surface surveying into capture from an onboard system and automatic labeling of defects according to a single industry classifier.
The type, severity, and coordinates of a defect go into one dataset that is comparable from survey to survey.
What changes after switching to machine control
Switching to machine control changes not only the method of capture, but also the unit of work planning.
Previously, surveying was selective, limited to the sections a crew could cover; now a single pass yields data for the entire route, and the repair decision rests on the actual picture rather than on requests from below.
Coverage and frequency
Survey speed stops depending on the number of people in a shift: the data stream is processed automatically, and an engineer steps in where the algorithm flagged a disputed situation.
This makes it possible to keep road network coverage going not as isolated surveys but as a regular cycle — the same road comes into view at the same interval, and the dynamics of deterioration is visible in comparison rather than from memory.
The number of repeat surveys drops for the same reason. If the initial pass tied a defect to a kilometer and lane, a follow-up check is needed only to confirm the class, and disputed cases are filtered out before the crew goes out on the road.
A single scale for assessing defects
Reproducibility of assessment used to rest on the experience of a specific engineer: two crews could assign different classes to the same crack.
Machine control sets a formalized attribute — geometry, extent, damage area — and assigns a defect to a class by a threshold that is identical for all sections and shifts.
A repeat pass over the same spot yields the same result, and divergence becomes not the norm but a signal of a problem in the data.
Report preparation time shrinks to processing time: the summary is assembled immediately, not after manual consolidation of tables. The result is a manageable cycle where coverage is known, the defect class is clear, and the scope of work is justified.
Surface capture and processing: options and limitations
The carrier determines not only the angle but also the class of defects that will even make it into the frame.
Below are four schemes for road surface capture that are realistically chosen between at project start; they can be combined within a single loop, but each adds its own stage of tying frames to coordinates.
| Option | When it fits | Limitations |
|---|---|---|
| Capture from a car | Highways, long linear sections, regular patrols | Depends on speed and lighting, ruts are partly hidden by traffic |
| Drone flight (aircraft of the DJI level) | Bridges, interchanges, rings, inaccessible shoulders | Flight regulations, motion blur in wind, short operating time |
| Walking survey with a handheld device | Sidewalks, local defects, disputed spots | Low productivity, subjective choice of points |
| Stationary camera on a section | Monitoring the development of an already found defect | Narrow field of view, requires power and a communication channel |
The key limitation of capture is not sensor resolution but conditions: glare from wet asphalt, shadows from fences, and twilight produce false positives more often than real wear. That is why we build repeat passes and calibration against reference sections into the plan.
Where to process: onboard or in the shared loop
Onboard processing gives a fast response: frames are labeled on site, and only the result goes out — the communication channel and storage are offloaded. The price is the limited compute resource and more complicated model updates across a distributed fleet.
The shared loop removes these limitations: the model is retrained on the entire accumulated history, an engineer rechecks disputed frames, and a full archive remains for relabeling.
A mixed scheme — preliminary selection onboard, final decision in the loop — usually gives the best balance, and that is what we adopt as the primary one.
How implementation proceeds: from section audit to handover
Implementation proceeds in stages, and each one closes with an acceptance criterion agreed in advance. That way the model enters the working loop not on someone's word: at every step it is visible what exactly was verified and on what data.
- Audit of sections and capture conditions. We drive typical stretches and record the surface, lighting, weather, angle, and capture speed. The output is a map of sections and a list of conditions under which the system is required to work.
- Defect class scheme. We agree on the list: longitudinal and transverse cracks, crack network, potholes, rutting, scaling, patches. For each class, written boundaries define what counts as a defect and what counts as background.
- Data preparation and labeling. We remove duplicates and blurred frames, balance rare classes, and label disputed frames twice and cross-check. Some sections are set aside right away — they will be used only for validation.
- Model training and calibration. We tune the confidence threshold so that thin cracks are not lost and false positives on shadows and puddles do not occur; stitching neighboring frames removes duplicates of the same defect.
- Validation on a hold-out set. We calculate recall and precision for each class on sections the model has not seen, plus a manual run in difficult conditions: rain, oncoming sun, night capture.
- Handover and training of your team. We deliver the model, the relabeling regulation, and the retraining instruction, and we review errors on your examples.
Acceptance criteria at each step
Criteria are fixed before the start: for the audit — a map of sections; for labeling — the share of agreement between labelers; for the model — metrics on the hold-out set separately for each class; for handover — your team runs the relabeling and retraining cycle on its own. A stage without a closed criterion does not move forward.
What we hand over: labeling, models, regulations
We hand over not a running service on our side, but the full loop that your team can maintain and develop without us.
The package includes everything for retraining, acceptance, and integration with accounting systems — from labeled data to the regulations by which operators capture and submit material.
- A labeled dataset: frames by defect class — cracks, potholes, rutting, edge failure — balanced by surface type and lighting conditions. A separate validation sample shows real recall, not reported recall.
- A trained model and the rules for updating it: which version is in operation, what data it was trained on, at which metrics a new version is admitted to launch, and how to roll back to the previous one.
- A scheme for integration with your accounting system: how defects are tied to sections and kilometers, in what form they go into reporting, and who confirms the record.
- A capture regulation: angle and frame overlap, permissible driving speed, conditions under which material is not counted, and the calibration procedure before departure.
- A report acceptance regulation: the miss rate at which a batch goes for recapture, the procedure for reviewing disputed cases, and response deadlines.
- Instructions for operators: checking material quality on site, actions for a blurred frame or a change in weather, and whom to report a failure to.
- Documentation and knowledge transfer: a description of the loop, a model change log, a review session with your team, and a support period after launch.
Rights to the model and data
We separately fix the rights to the model and data in the contract: the labeled dataset, the model weights, and derivatives remain yours. You can retrain it with your own resources or hand it to another team without losing functionality. What remains with us are only anonymized methodologies that do not contain your data.
Case study: highway surveying before and after implementation
The section is 18 km of an urban highway with six lanes, overpasses, and shaded courtyard entrances. Three crews surveyed it for about two weeks: defect coordinates were refined on site, and disputed remarks were closed with a repeat survey.
By the time the report was released, the service was already fixing some of the defects, and the picture diverged from the facts.
We moved capture to panoramic cameras with geotagged frames, and reduced classification by type and criticality to the requirements of GOST R 50597.
The model is trained on labeling by road engineers: frames with low confidence go to human confirmation rather than into the report.
- Defect detection recall — 62% → 96% on a control sample of 1,200 manually labeled frames.
- Report preparation time — 9–12 days → one business day after the survey: the defect map and the quantity schedule are generated automatically.
- Repeat surveys — 4 out of 10 sections → 1 out of 15, and only where the view was blocked by traffic or a tunnel.
What exactly delivered the gain on difficult sections
On shaded and wet surfaces, the gain came not from the model itself but from a combination of solutions.
Frame stabilization at traffic speed together with geotagging removed duplicates of the same defect from different passes, and a second pass over low-confidence frames cut the false positive rate from 18% to 4%.
Courtyard entrances and marking seams were relabeled manually by an engineer — those frames went into retraining, and the next survey ran cleaner.
How does neural network inspection differ from threshold filtering?
Threshold filtering is a set of rules like "a region of the frame is darker than the background by a given amount → defect". It works predictably and requires almost no data, but it rests on an assumption: a defect always contrasts with the surface in the same way. As soon as the assumption breaks, the detector breaks too.
And it breaks regularly. Morning sun casts shadows from poles and trees that pass the threshold and become cracks; wet asphalt merges in brightness and hides potholes.
The change of season adds snow, de-icing agents, and markings — each factor needs its own set of thresholds, and they are reconfigured manually.
Neural network inspection learns from labeled examples: engineers show the model frames with and without defects — from different cameras, at different times of day, across seasons and surface types.
The model relies not on brightness but on structure: the geometry of a crack, graininess, edge rupture. Robustness to lighting comes from the fact that the sample already included shadows, blown-out highlights, and wet surfaces.
False positives in threshold filtering grow along with sensitivity: lower the threshold and you catch more cracks and at the same time more stains, seams, and traces of sealant.
A neural network separates these classes, and their level is regulated by the confidence threshold and by error review: engineers look at what exactly the model took for a defect and send such frames back into training.
The practical conclusion: thresholds remain a fast filter for cutting out obviously empty frames, while the decision about a defect is made by the neural network — robustness to lighting and season no longer requires rewriting rules for every section.
Common technical doubts before project start
Doubts at the evaluation stage usually concern not the model but capture conditions and the boundaries of the loop: what to do about rain and twilight, how to deal with winter surfaces, and whether it is possible to avoid exporting data outside. Below are the answers in the form in which we give them to the customer's engineering team.
How is capture carried out in rain and twilight?
Capture in bad weather and night surveying is not a separate mode but an extension of conditions in the training sample: droplets on the optics, headlight glare, reflections from wet asphalt.
Frames below the sharpness threshold are not discarded — they are flagged as uncertain and go to manual review.
How does the model distinguish a pothole from a shadow and a repair trace?
In winter a pothole is often covered by ice crust, and the shadow of a snowbank resembles a depression; a repair trace looks smooth and geometrically repetitive.
We separate them by the stability of the contour between neighboring frames: a shadow trembles, a defect holds its shape, and water and dirt accumulate next to a pothole.
Data must remain within our loop — is that possible?
Yes, no external export is required. The loop is closed: capture, labeling, training, and model operation are deployed on agreed customer capacity or in an isolated segment. Weights, logs, and labeling stay inside, and access is segregated by role.
How often should the model be retrained?
Retraining is triggered by an event, not by the calendar: labeling has accumulated for new defect types, the capture carrier has changed, the surface has moved into a different season.
Scheduled runs are usually tied to the change of season, while unscheduled runs occur if the share of uncertain frames has grown.
Let's discuss the task: how the scope of work is estimated
The scope of work for AI road inspection is estimated not by a price per kilometer but by an inventory of the customer's data and processes.
Below is the list of what removes uncertainty fastest: from these answers we assemble the work plan, fix the acceptance metrics for results, and set the boundaries of responsibility.
- Register of sections: length, road class, surface type, number of lanes, bridges and overpasses — the volume of reference labeling and validation depends on this.
- Capture conditions and frequency: what we capture with (panoramic camera, dashcam, mobile survey), time of day, season, weather, frequency of passes — this sets the requirements for model robustness on shadows, wet asphalt, and worn markings.
- Reporting requirements: which defect classes we recognize (longitudinal and transverse cracks, crack network, pits, rutting), size thresholds, export formats, and coordinate referencing.
- Accounting systems: the existing GIS, defect register, road passport, and the escalation procedure for repair requests — this determines which integration layer is needed.
- Those responsible for accepting results: who signs off, by which metric (detection recall and precision on a control sample), and how a disputed case is reviewed.
- Access and data: the archive of video and photos, historical passes from previous years, and dedicated reference sections for calibration.
After that, what remains is to agree on the work plan: a pilot section, metrics measured on it and cross-checked against manual walking, then scaling to the entire register on a clear schedule. Describe the task in these points — and we will come back with an estimate and a plan.
