A SMALL PROJECT. A SMARTER PERSPECTIVE.

Built to understand
the spaces between.

SmartPark AI is a free, non-commercial learning project exploring a familiar problem: knowing where to park.

From parking images to useful insight.

Select an image, outline parking spaces, and run real vehicle detection. The app estimates occupancy using the intersection between detected vehicle boxes and your parking polygons. It does not infer occupancy from a box center alone.

Built for the browser and Cloudflare.

Next.js, TypeScript, Tailwind CSS and editable shadcn/ui components power a static frontend prepared for Cloudflare Pages Free. A separate Hono API runs on Cloudflare Workers Free for geometry validation. ONNX Runtime Web runs the lightweight YOLOX-Nano model on your device using WASM CPU inference. The model is pretrained on COCO and uses the published Apache-2.0 license. Turso is reserved for future genuine aggregate results; current routes do not write to a database.

Privacy belongs in the architecture.

Image pixels never leave your browser. The model and runtime load from this site; no paid inference service or GPU is used. Only polygon coordinates and dimensions are sent to the geometry API. Local preview and actual results remain in app memory until cleared or the page is refreshed. Optional session saving contains geometry and an image fingerprint, never image bytes.

What comes next.

Stage 3 includes real browser detection, bounding-box overlays, configurable thresholds, occupancy calculations and an actual-results dashboard. Future work includes improving dense-lot accuracy and optional aggregate-only persistence. Video processing is not implemented.

Understand the limits.

A small 416-pixel model can miss distant or occluded vehicles and can confuse vehicle types. Parking occupancy is a geometric estimate, not a guarantee that a space is safe or available. Always verify the image visually.

Follow the project.

Project repository and contact details have not been configured yet. The footer links to related public GitHub resources. Model and sample attribution files ship with the site.