Street Parking Presence Inference from Street-Level Imagery

This project studies street parking presence inference from street-level imagery. The core idea is to detect image-level parking-related cues such as parking signs, parking meters, and curb-related evidence, and then aggregate these cues across nearby views to infer whether a street segment supports on-street parking. The framing is motivated by a practical reality of the task: parking cues are often small, sparse, occluded, and viewpoint-dependent, so single-image predictions are brittle — a negative result from one image may simply mean the relevant cue is outside the field of view or too small to detect reliably.

We focus on three visual cues:

• Parking signs, modeled as supervised object detection on the Mapillary Traffic Sign Dataset (MTSD).
• Parking meters, evaluated in a zero-shot cross-dataset setting using a COCO-pretrained detector on Mapillary Vistas.
• Curb structure and curb color, modeled through binary curb segmentation followed by a heuristic HSV color analysis.

Among these, the parking-sign detector emerges as the strongest explicit cue. Parking meters provide some transferable signal but are noisy. Curb color is visually meaningful in some cases, but it is sparse and difficult to recover robustly — especially when segmentation masks are thin or contaminated by nearby road markings.

The final project result is not just a set of independent detectors. The detectors are used as image-level components inside a segment-level aggregation system. This is important because the true downstream task is closer to "does this local street segment contain parking-related evidence?" rather than "does this single image contain a visible sign?". The aggregation experiments show that pooling evidence across multiple views substantially improves recall compared with single-image inference, while retaining useful precision.

The overall result is a multi-cue system in which:

• parking signs provide the strongest explicit parking evidence,
• parking meters provide a weak but non-random auxiliary signal,
• curb segmentation is feasible,
• curb color is possible to estimate conservatively, but should be treated as a sparse auxiliary cue rather than a primary standalone signal,
• segment-level aggregation substantially improves recovery of parking-related evidence when cues are sparse or only visible from some viewpoints.