Teaching a camera to spot debris before a human could

Radar keeps track of the large objects in orbit, but the small ones slip through. A fragment the size of a pen can go undetected and still strike the International Space Station hard enough to matter.

The constraint was never accuracy in a lab. It was accuracy at frame rate, on hardware that could survive the environment, picking a small fast object out of a noisy moving frame in the time it takes to be useful.

We built a detection and tracking pipeline on NVIDIA DeepStream, wired into the ISS's external cameras and running inference at the edge rather than shipping video to the ground. The system scans the surrounding space continuously and flags objects radar would miss.

On detection it analyses the object's trajectory, speed and size and estimates the likelihood of a collision. When that likelihood is high it alerts mission control automatically, with the detail they need to act, rather than surfacing raw video on a screen no one is watching.

Mission control gets an early, specific warning about the debris radar cannot see, in time to modify the station's orbit or nudge the object onto a re-entry path where it burns up. Detection, tracking and alerting run as one pipeline, at the edge.

The work put production computer vision into one of the least forgiving environments there is, and set the pattern for everything that followed: run the model where the data is, and land the answer inside the workflow that already exists.