Road Asphalt Monitoring with Computer Vision
Continuous CV monitoring of asphalt condition replaces periodic manual surveys. The data pipeline, the cost reality, and where it earns its place.
Pavement condition surveys used to be the slow-and-expensive part of asset management. Specialized survey vehicles ran routes every 2–3 years, manual review followed, the data fed into the pavement management system. The 2026 alternative: continuous capture from connected vehicles, processed by AI, feeding the same systems with much fresher data.
Where this earns its place — and the realities.
How it works#
Cameras (sometimes specialized survey vehicles, increasingly regular DOT vehicles or even crowdsourced) capture continuous video of the pavement surface. ML models trained on pavement-defect taxonomies (alligator cracking, rutting, raveling, patching) classify and locate defects.
Output: per-segment, per-month condition data instead of per-segment, per-three-years. The freshness change is the value.
The vendors and approaches#
Specialized vehicle providers. Roadway Asset Services, Mandli, Pavemetrics. Highest accuracy; highest cost per mile. The traditional approach.
Connected-vehicle / fleet integrations. RoadBotics (acquired by Michelin), Pavementscan, plus crowdsourced approaches. Lower per-mile cost; somewhat lower precision but much fresher.
In-house pipelines. Some larger DOTs (Texas DOT, Caltrans, Florida DOT) have built internal capabilities.
Choose by network size, freshness requirement, and existing data infrastructure.
Where it earns its place#
Large urban networks where conditions change quickly and the standard 2–3 year survey cycle is too slow.
Toll and managed-lane operators where condition is part of the customer experience.
Programs aiming for ISO 55000-style asset management. The data freshness is a prerequisite.
Mixed-quality networks where targeted intervention beats blanket treatment.
Where it doesn’t#
Rural networks with low traffic. Standard survey cycles work; continuous capture is overkill.
Programs without capacity to act on findings. Faster data without faster decisions is overhead.
Networks with poor base data. AI deteriorates without ground-truth calibration.
The data-pipeline reality#
The pipeline runs:
- Vehicle capture (specialized or integrated)
- Upload to cloud (often the bottleneck for high-resolution capture)
- AI processing (defect detection, classification, location)
- QC sampling
- Update to pavement management system (PMS)
- Decision-support analytics
Each stage has integration complexity. Our data engineering practice builds out this pipeline for road owners.
The accuracy reality#
Continuous-capture AI tools typically achieve 85–92% agreement with manual surveys on standard defect types. The disagreement is mostly on borderline severity calls and unusual defect types.
The discipline that makes this work:
- QC sample 1–5% of segments manually each cycle
- Compare AI vs manual; investigate disagreements
- Refine the model with the firm-specific calibration
What we ship for road-asset programs#
For continuous pavement monitoring engagements:
- Capture-to-PMS pipeline
- QC workflow with manual sampling
- Network-level analytics dashboards
- Treatment-decision support integration
- Audit trail at every AI-involved step
The economics#
For a network of ~1,000 lane-miles, continuous AI-based capture costs $50k–$200k/year all-in. The traditional 2–3 year specialized survey costs roughly the same in amortized annual terms but produces stale data.
The freshness advantage compounds in capital planning: better-timed interventions save more than the cost differential, even before the data-quality improvement.
Continuous CV pavement monitoring earns its place at scale. The pipeline is the work. Our team builds the data layer for road asset programs. Tell us about the network.