ParkingSpotter

• 

  Notification email for open parking on Wythe Ave

Inspiration

Finding street parking in NYC is purified intravenous cortisol. Finding a spot is one thing, translating the zone restrictions and making sure you don't get a ticket is another. Just to get around, I feel like I needed an expert in NYC parking laws with a real time view of the city streets.

What it does

A custom trained AI vision model watches live feeds from NYC DOT traffic cameras and alerts users when a spot on a given street becomes available.

How we built it

From a Next.js frontend served from Cloudflare, users can browse available cameras in NYC and submit their email under a camera / street they want watched. Once submitted, the frontend calls out to my container on fly.io to enter the user's email into my DB under the camera they specified. A python process from the same container polls the DB for cameras with watch requests, pulls the latest image from that camera feed, passes it to my model for inference and parses the output to adjust the status of the street accordingly, i.e. parking or no parking. Emails are sent with live feed and bounding boxes around the spot when the status changes.

Challenges we ran into

Training model on empty space: How do you teach something to recognize nothing? Once I have it trained, how do I design a system around this raw classifier? How can I abstract the inference away from the user?

Accomplishments that we're proud of

It works. I built everything myself from annotating the dataset to deploying the production website.

What we learned

When the object you're trying to detect itself is not detectable, you have to teach the model what surrounds the object. For this I need to draw irregular polygon bounding boxes in the training data.

What's next for ParkingSpotter

I want to include a link in my alert email to a user feedback process that will improve model. Annotating more data for a larger dataset myself will also help. I can also use other NYC APIs to read NYC street zones and build a chat agent to decode NYC's byzantine street parking laws to non New Yorkers. I'm also interested getting more cameras. Out of the 900 available on the NYC DOT website, only about 150 are pointed towards valuable parking. I could do some installs like damnlines.com did for recording lines in front of popular restaurants with long lines. I could ask property owners to install them or use public poles and install them myself. If users are worried that a spot will fill up between then time they get notified to when they drive to the spot, parkingspotter.nyc will deploy an autonomous drone to hover over that spot dangling a traffic cone underneath it and a LED sign of the users email.

In addition to improving the core functionality, I will also be building a chat agent that can view NYC parking zones and help decode them for non New Yorkers. I plan to run statistical analysis on historical parking patterns in the feeds and let my agent query this custom database. With this, parkingspotter.nyc users will have an all-seeing NYC street parking expert in their pockets.

Built With

• cloudflare
• docker
• fastapi
• next.js
• pytorch
• yolo