Self-supervised obstacle detection during autonomous UAS taxi operations

Abstract

This research explores the application of self-supervised learning techniques for obstacle detection and collision avoidance during UAS auto-taxi. Autoencoders were used to detect obstacles as anomalies by comparison of reconstruction errors. RGB cameras and millimetre wave radars covering conflict free zones (CFZs) around the own-ship were chosen to provide inputs to autoencoders. Results demonstrated that autoencoders were able to detect obstacles as anomalies within the CFZs but with certain limitations at lay the foundations of further work and investigation within the research area.