New AI Traffic System Watches 400 Cameras at Once

Researchers at the Indian Institute of Science have built something called TrafficFab — a test rig for AI-powered traffic management in giant cities. The problem it tackles is simple to state and hard to solve: a megacity like Bangalore has thousands of CCTV cameras, and somebody has to watch all of them, in real time, without drowning in video data or burning a fortune on electricity and cloud bills.

Their answer is to split the work. Small, cheap computers — Raspberry Pis and Jetson chips, the same kind of hardware used in hobby robots — sit near the cameras and do the first pass of analysis right there. Heavier number-crunching, like forecasting where traffic will snarl in the next 15 minutes, happens in the cloud. The system automatically shifts work around depending on power costs and how busy things are, like a restaurant manager moving staff between the kitchen and the floor as the dinner rush builds.

The clever part is the learning. Instead of shipping everyone's video to one place to train the AI, the models improve on the local devices and only share what they learned — not the footage itself. That means your license plate and your face don't have to travel across the city to make the system smarter. In tests, TrafficFab sustained real-time analysis of about 400 live camera streams, roughly 10% of Bangalore's network, across a mix of hardware.

So what's the catch? This is a lab testbed, not a deployed city system. It was evaluated on a Bangalore-inspired simulation, not on live public roads, and it needs a city willing to install and maintain hundreds of small devices on street corners. Real megacity rollouts also raise questions about who watches the watchers.