Sensing, signal processing, and navigation

Small unmanned aircraft depend on GPS, and GPS can be jammed, spoofed, or simply blocked indoors and in dense terrain. This project set out to give a small drone a reliable estimate of its own position without it, and without prior knowledge of its starting location, landmarks, or the signals around it.

Three independent sources

Signals of opportunity. Five phase-coherent software-defined radios, all driven from a single clock and paired with five omnidirectional antennas, measured the direction of radio signals already present in the environment. Correlative interferometry and the MUSIC algorithm turned those measurements into bearings for triangulation.

Visual odometry. A stereo camera and NVIDIA embedded computing estimated the drone’s motion from sequential images in real time: image capture, SIFT feature extraction, brute-force feature matching, then motion estimation in X, Y, and Z.

Inertial and environmental sensing. Onboard motion and environmental sensors provided a continuous, high-rate estimate between the slower radio and camera updates.

Fusion

An error-state extended Kalman filter fused the visual odometry and inertial data, and a neural network learned the harder, non-linear relationships between all three sources. Training data came from an RTK GPS receiver on the test platform, which recorded ground truth position during test runs. Combining the sources raised the signal-to-noise ratio and reduced the uncertainty and ambiguity that any single sensor carried on its own.