AI-Powered Infrastructure Monitoring from Municipal Dashcam Fleets

The county's public works department relied on manual inspection to track infrastructure deterioration — inspectors drove periodic routes and logged issues by hand, creating a maintenance backlog that was perpetually weeks behind reality. With over 200 lane-miles of roads, dozens of intersections, and thousands of signs to monitor, the traditional approach couldn't keep pace. The county wanted to mount dashcams on existing municipal vehicles and use computer vision to flag issues automatically from the video feed. But that required training data: tens of hours of real local footage with labeled infrastructure defects, captured under the specific conditions those vehicles encounter — wet roads, overcast skies, night patrols, residential streets alongside commercial corridors.

We deployed a fleet of vehicle-mounted dashcams across the county over a focused collection campaign, capturing 75+ hours of street-level footage covering every major road, intersection, and residential block. Routes were planned for geographic completeness — ensuring diverse conditions including residential neighborhoods, commercial corridors, school zones, and industrial areas under varying weather and lighting.

Annotation combined automated AI enrichment with expert human review. First, a vision model analyzed representative frames from each clip to generate structured infrastructure condition reports and bounding box detections across 10 defect categories. Human annotators then validated and corrected every detection — verifying bounding box accuracy, reclassifying edge cases (a patched crack vs. active deterioration), and adding defects the model missed. This hybrid approach delivered the precision of human judgment at the speed of automated processing.

The detection taxonomy was co-developed with the county's public works team to match their existing maintenance categorization system — ensuring that model outputs map directly to work order types. Each clip's final annotation combined machine-generated bounding boxes, human-verified labels, natural language scene descriptions, and structured metadata — creating training pairs rich enough for the county's model to learn from.

The hybrid pipeline processed the full 75+ hour corpus in days rather than the months that purely manual annotation would have required, while maintaining the accuracy standard that safety-critical infrastructure monitoring demands.

The enriched dataset became the training foundation for the county's automated infrastructure monitoring system. Municipal vehicles now capture footage as part of their normal routes — garbage trucks, utility vans, and patrol cars all contribute to a continuous street-level survey without any additional cost or staffing. The model flags defects in near real-time, automatically generating maintenance tickets categorized by severity and location. In the first month of deployment, the system identified 3x more infrastructure defects than the previous manual inspection process — the majority being cracked road surfaces and faded lane markings that would have gone unnoticed until they became safety hazards. The county estimates the automated system covers every road segment at least twice per week, compared to the previous quarterly manual inspection cycle.