ThinkSwiss · Bern University of Applied Sciences · Jun. — Aug. 2023
No-GPS Drone Navigation
Vision-based drone navigation for GPS-denied environments, including a redesigned segmentation network and a synchronized onboard sensor pipeline.
Research in the loop
During my ThinkSwiss research stay at Bern University of Applied Sciences, I worked on a drone designed to navigate through GPS-denied environments using onboard perception and sensing.
Perception on constrained hardware
I redesigned the segmentation network around a transformer decoder and combined augmentation with transfer learning. The resulting model improved F1 score from 64.11 to 88.38 while sustaining 67 FPS on a Jetson Orin Nano.
I also led selection and integration of cameras, altimeters, and inertial sensors, then developed a synchronized acquisition pipeline with filtering and augmentation so downstream components received reliable data.
Beyond the benchmark
This was as much a systems project as a model-training project: every accuracy improvement had to survive an embedded compute budget, real sensors, and flight constraints. The linked demo and field notes show that wider context.