Abstract Manager

Urban curbside parking availability is constrained not only by parked vehicles and regulatory signage, but also by physical access, such as driveways and alleys, that interrupt usable curb length and impose parking-clearance requirements. However, these access features are rarely mapped at scale in residential curbside inventories because they are irregular, frontage-specific, and often affected by parked vehicles, vegetation, and oblique camera perspectives.

This paper presents and evaluates a mobile vision-GPS/GNSS framework for generating a verified inventory of curbside access features in urban areas. Results showed that fine-grained driveway labels were visually unstable under mobile video conditions, while a consolidated Narrow/Wide access structure improved deployment suitability. The selected YOLO26-XL model achieved a mask precision of 0.820, a mask recall of 0.787, and an inference latency of 18.36 ms/frame.

A multimodal LLM verification layer was introduced to confirm object presence, correct narrow/wide label errors, and flag unusual cases for review, with Gemini 3 Flash providing the strongest multi-criteria trade-off across verification performance, response time, and per-detection cost. Prompt tuning further reduced the aggregate unusual rate from 0.162 to 0.079, achieving an aggregate F1-score of 0.931 and accuracy of 0.901 across 699 candidates.

The final output is a GIS-ready, verified inventory of curbside access constraints that can support more realistic estimates of effective residential parking supply and future curbside management analyses. The proposed framework combines computer vision, instance segmentation, multimodal LLM verification, and GPS/GNSS alignment to support the detection of curbside infrastructure.

Keywords: Curbside infrastructure; Parking utilization; Instance segmentation; Computer vision; Driveway and alley detection; Residential Parking Program.