Drone traffic

First time working with vision-language models. The brief was to turn raw drone footage into something that understands traffic. I finished 3rd at the FlytBase Visual Intelligence Hackathon.

Problem

A drone sees the road. The footage does not say which vehicle is which, which lane it is in, or what the traffic is doing. Everyone at the hackathon got the same footage.

What I built

A dashboard that detects and tracks cars, buses, trucks, motorcycles, and pedestrians, and keeps their identities through occlusions and crossing paths. Each one gets a class, a real-world speed, and acceleration.

The tracks roll up into the road: volume and direction, which lanes are busy, speed profiles, queue length, density, occupancy, and how flow moves across it.

The footage included SRT telemetry: GPS, altitude, gimbal orientation, and timestamps. That puts a vehicle on a road, in a lane, heading a direction, at a speed.

Result

300+ people registered. 70+ built it in the room. 3rd place.

Post