[aunty_helen]

Hey this is something I've been working on as a side project for the last 2 years. Using traffic cameras to detect the flows through intersections with the goal of providing better stats to governments and city planners.

Hopefully I can have the poc going in the next month. A city wide integration with something like AB street would be really cool and was phase 2 for this project. Phase 3 being an optimised citywide traffic control.

[dabreegster]

Whoa, this would be awesome! Strongly consider publishing open datasets. :) Could you talk at all about how it works, or what privacy issues you've dealt with?

[aunty_helen]

Sure, the system takes a raw camera feed and spits out an entry into a sql db which will then, depending on customer requirements, be provided in a format they want.

In between these steps, (all based in python) opencv takes the video, splits it, sticks some of the frames on a kafka cluster.

Another node picks up approx 2 fps, hands it to a object detection algo based on tensorflow which then puts the results on another kafka topic.

A heavily customised tracking algo pulls all of the frames available and the tensorflow results for initialisation and reverification. Once a vehicle is tracked through the field of view the vehicle is counted.

For the privacy issues, as this is only tracking for a short amount of time, not tracking between intersections and not tracking persistent uuids or numberplates, there shouldn't be too much trouble. But it will depend on the local law requirements. As this is a statistical problem, tracking the actual path of a car between point a and b isn't completely necessary if you can get a sufficient amount of coverage over a city. (You probably already understand this.)

As for opening the datasets up, I plan to do this with the poc. I've been using the Biloxi Mississippi camera live stream from youtube for all of my testing so far. My poc will be a website with this live stream having an overlayed view and the ability to download the stats for the day/month/year etc. As for other cities, I'm sure a FOIA will be able to get you the data. But I will be pushing to publish it openly as they will have to pay for it either way.

As you can probably tell, I'm trying to turn this into a business and the costs of actually collecting and processing the data isn't something that is reasonable for single person to bear. After talking to the traffic dept for my country, the largest city of 1.5m has about 3000 cameras. But I am aiming for a system that's cheap enough to enable 24/7 monitoring as that's when this all gets pretty exciting.