All projects
2021 Archived

Search-and-Rescue Quadcopter

YOLOv3 obstacle avoidance for alpine rescue

  • YOLOv3 tiny, running on-board
  • Best in CS Vienna International Science Fair
  • A→B autonomous point-to-point flight

The project that got me into computer vision, and the one that won Best in Computer Science at the Vienna International Science Fair.

In alpine search-and-rescue the constraint is time. A drone that flies itself from A to B without a pilot, and without hitting anything, covers ground a human team cannot.

A YOLOv3-tiny network locates obstacles in the drone’s field of view during flight, and a custom avoidance algorithm built on trigonometry computes the shortest path around a detection rather than stopping or reversing.

Choosing the smaller, less accurate model was the decision that mattered. A better detector that produces its answer after you have hit the tree is worth nothing. First time I ran into the latency-versus-accuracy trade-off that has shown up in every vision system I have built since.