Ground-Based Lightweight Machine Learning Model for GPS- Denied Navigation in Uavs

1.   Dasari Avinash
2.   Prasad Ghumare
3.   Aryan Hajare
4.   Puppala Santhosh
5.   Sachin Deshmuk
1,2,3,4 B. Tech (AIML)student, Department of Computer Science and Engineering, Sandip University, Maharashtra, India.

5 Professor, Department of Computer Science and Engineering, Sandip University, Maharashtra, India.

Abstract - Flying small unmanned aerial vehicles (UAVs) in outdoor areas without GPS usually means strapping a heavy, power-hungry companion computer to the drone to handle visual navigation. For budget-friendly research setups under$700, this simply isn't practical. This paper explores a more cost-effective alternative: shifting all the heavy visual processing to a regular consumer laptop down on the ground. The drone itself only carries its flight controller, a basic camera, and a digital telemetry link. By pairing a MicroAir H7 flight controller (running PX4 v1.14) with a SIYI HM30 data link, we pass video to a compact machine learning model on the laptop. The system then sends pose estimates back to the drone at 18 Hz. Because the radio link introduces a 178 ms round-trip delay, the ML model has to be extremely fast—which is exactly why we kept the architecture so lightweight. During 240-meter outdoor test flights, the drone successfully held its position with an absolute trajectory error of less than 3.4%, all without draining the aircraft's battery for computing tasks.Key Words : GPS-denied navigation, off-board processing, PX4 autopilot, MAVLink, visual-inertial odometry, budget UAVs.