Smart mapping of urban roads: the municipal fleet as a pavement sensor

Technical note.

Summary. Road maintenance in Brazil is still reactive: a pothole becomes a work order only after the complaint — and after the damage. This article describes a "fleet as sensor" approach in which vehicles the municipality already operates (buses, garbage trucks, maintenance vehicles) continuously capture the pavement; a computer vision model detects and classifies anomalies; each occurrence is geo-referenced with photo, date and severity; and a central dashboard prioritizes repairs by criticality, traffic and cost. The result is a continuous, auditable inspection with no new construction — from detection to proof of repair.

Potholes are today one of the biggest sources of avoidable cost in urban management: vehicle damage that turns into court-ordered compensation, emergency repairs up to 3× more expensive than planned maintenance, and public perception that deteriorates with every rainfall. The current model — manual inspection and citizen complaints — is slow, depends on the damage already existing and generates no data for prioritization.

• Reports arrive by phone/ombudsman channels, without precise location or evidence.
• Without severity and traffic data, crews respond in the order complaints arrive — not where the risk is greatest.
• There is no auditable trail: where a pothole was, when it was patched, with what quality — information essential to oversee pothole-repair contracts.

The system operates across four layers, requiring no new construction and without interrupting fleet operations:

• Capture — compact in-vehicle cameras on the municipal fleet. Each vehicle covers its normal route; the network is revisited continuously, with no added travel cost.
• AI detection — a computer vision model (YOLO family / segmentation) trained for Brazilian pavement, classifying potholes, cracks, deteriorated patches and depressions, with dimension and severity estimates. Runs on the edge (embedded) or in the cloud.
• Geo-referenced data — each detection records photo, GPS coordinate, road, date/time and severity in a geospatial database (PostgreSQL + PostGIS). Repeated detections of the same point are consolidated into a single occurrence with history — including visual proof of the repair.
• Management dashboard — a city map with occurrences by severity, prioritization (criticality × traffic × recurrence), repair-crew SLAs and executive reports.

Each layer uses reference technology that is mature and portable:

• Capture — IP67 automotive camera with GNSS, offloading via 4G/Wi-Fi.
• Inference — YOLO/segmentation on the edge (Jetson or similar) or in the cloud, depending on connectivity and cost.
• Data — PostgreSQL + PostGIS, with deduplication by radius and by road.
• Services — REST API and webhooks for integration with 156/ombudsman systems and public-works ERP.
• Visualization — a web dashboard with the network map and BI (Metabase).
• Security — automatic anonymization (blur) of license plates and faces; LGPD by design.

The entire platform runs in the cloud or on the municipality's own infrastructure, according to local data policy — the data belongs to the municipality, in an open format, from the first day to the last.

The system observes the pavement, not people. Vehicle license plates and faces are automatically anonymized (blurred) before storage; the images record exclusively the public asset (the road). Processing is grounded in the exercise of public policy (art. 7, III and art. 23 of the LGPD), with an impact assessment (RIPD) delivered at implementation and an access audit trail.

The same infrastructure — cameras on the fleet + a computer vision pipeline — detects far more than potholes, through a model update: missing manhole covers, flooding, illegal trash dumping, dark streetlights, damaged signage and fallen trees. It is the natural path toward a smart-city platform, without swapping out the hardware. The "fleet as sensor" approach is already operating in Brazil — Manaus, São José dos Campos and Sabesp use variations of the same idea — which lowers the innovation risk for the municipality.

We integrate SASCAR, Autotrac, Omnilink and similar platforms day to day: installing and operating embedded sensors is our home turf for more than two decades. We build and operate our own platforms (WMS, TMS, BI, data enrichment) — we don't resell black boxes. And we design dashboards for public decision-making: prioritization, SLA and accountability, with established practice in LGPD, pentesting and data governance. For the city government, it starts with a free 48-hour assessment: a survey of the available fleet, routes and systems to integrate, and a network-coverage plan with a 90-day pilot.