From Dashcam to Decision — Automating Pavement Condition Surveys

Road maintenance has two immovable constraints: the network is vast, and the condition data is stale the moment you collect it. Every kilometre that sits between one inspection cycle and the next is a kilometre where distresses are progressing, water is intruding, and the cost of eventual repair is compounding. The only question is how quickly you can close the loop.

Across four separate engagements — a Network Survey Vehicle analytics platform for a highway authority consultancy, a standalone desktop inspection tool for survey firms, a dashcam-driven field app, and a web platform for airport pavement teams — we've watched the same pattern repeat. The teams that compress survey-to-report turnaround from weeks to hours change what they can do strategically, not just tactically.

Why manual surveys are worse than they look

The conventional numbers understate the problem. A two-person crew walking a kilometre of highway can note surface defects for about four hours a day at modest accuracy. Weather, traffic, and fatigue eat another chunk. What reaches the office at the end of the day is clipboard notes that someone now has to reconcile against a chainage reference and type into a spreadsheet.

Three second-order costs compound on top:

• Inter-rater variability. Two inspectors walking the same kilometre will disagree on severity classifications roughly 20–30% of the time. At scale, that variance shows up as noise in trend analysis across successive surveys.
• Audit fragility. A clipboard note from eight weeks ago is not an evidence trail. When a repair contractor disputes a finding, you cannot re-examine the defect without going back to the site.
• Sampling, not coverage. Manual surveys inevitably sample. The regions that get inspected are the ones that are easy to reach, not the ones most likely to contain problems.

Each of these is solvable individually. None of them are solvable cheaply within a manual workflow.

The automated alternative

The pattern we've converged on across the four deployments looks roughly the same:

Capture. Drive-through video — from a survey vehicle, a maintenance truck, a dashcam, or a phone in a car mount. The footage has to carry enough metadata to be self-referencing: GPS, timestamp, chainage, vehicle speed. In field cameras that don't emit structured telemetry, we read the burnt-in banner with OCR.

Detect. A segmentation model localises distresses frame-by-frame. The inventory matters. On highway projects we classify into fourteen specification-aligned distress types; for simpler dashcam projects, two or three of the most commercially significant defects. Every detection gets a severity rating against configurable thresholds.

Deduplicate. This is the step that matters most and is usually done worst. A single crack visible in the camera for two seconds will appear in thirty to sixty consecutive frames. Naive systems count it thirty times. The fix is spatiotemporal — track detections across frames, bucket them by their geo-referenced chainage, and emit one record per physical defect regardless of how long the camera lingered.

Measure. Pixel dimensions aren't useful. Length in metres and area in square metres are. A perspective-aware conversion layer — using camera height, mounting angle, and the row of the detection in the frame — converts raw bounding boxes into physical units that map onto engineering specifications.

Report. The output of every deployment is the same shape: an annotated video, a deduplicated defect inventory, a segment-wise summary aggregated to some chainage interval, and a structured export that loads cleanly into GIS dashboards and asset-management systems.

The numbers, without hand-waving

One deployment surveyed roughly 900 km of state highway across twelve days of driving. The client's previous annual survey took sixteen person-weeks of manual review. Automated turnaround was under forty machine-hours — processed overnight on a single workstation, with one human reviewer verifying outputs the following day.

Another — an airport pavement platform — compressed a session that historically shipped a PDF two weeks after site visit into one that renders a full dashboard the same afternoon.

A third — a desktop inspection tool with a six-class distress taxonomy — let a two-person consultancy take on six simultaneous active surveys where they'd previously juggled two.

The specifics vary. The pattern doesn't.

What the tool cannot do

Automated survey doesn't replace engineering judgment. What it does is deliver that judgment to the engineer much earlier, with much better evidence attached. An experienced pavement engineer looking at an annotated 30-minute video clip overlaid with severity classifications makes better decisions, faster, than the same engineer with a clipboard and a spreadsheet. That's the real compression.

The software is the inexpensive part of the system. The expensive parts are the domain knowledge of the engineers using it and the operational commitment to re-survey often enough that the data stays fresh. Solve the inexpensive part first and the rest becomes a decision about how much coverage you want, not how much you can afford.