Type: multi-object tracking (MOT) · Input: sequential frames from a moving drone camera · Output: tracked trajectories with per-frame boxes and attributes.
We tracked vehicles across sequential frames, assigning each a unique track ID and maintaining it over time while classifying type — car, truck, bus, motorcycle, bicycle and more — and recording per-frame attributes.
Why tracking is hard
Following vehicles through video is harder than detecting them in a still: appearance shifts with lighting, shadow and weather; objects are occluded and re-emerge; and a moving drone camera changes the scene continuously. Holding a consistent identity through all of that is the core challenge.
What tracking reveals
Detection tells you what is present; tracking tells you what is happening. Continuous trajectories expose speeds, turning movements and flows — the dynamic layer that traffic studies, safety analysis and simulation depend on.
