Visual GPS-denied multi-agent localization & terrain classification

Abstract

This paper furthers the state of the art of visual localization by harnessing recently developed scene recognition and understanding machine learning (ML) algorithms to rethink how vision can be used for navigation, particularly when combined with a realistic, minimalist, and robust multi-agent approach. The focus on visual navigation comes as GPS is increasingly unreliable (jammed, spoofed, or disabled), or unavailable (indoor, or extraterrestrial environments). GPS denial/unreliability is an area of active research in general, with unmanned parafoil systems being one UAV application. The landing locations for parafoil systems must be pre-programmed manually with global coordinates, which may be inaccurate or outdated, and offer no in-flight adaptability. This paper will introduce a novel and minimalist approach to visual navigation and multi-agent communication using semantic machine learning classification and geometric constraints. This approach enables localization and landing site identification for multiple communicating UAV systems deployed in GPS-denied environments. This is a capability extending beyond singleagent localization, as well as a novel landing site classification approach. A mathematical foundation using information theory provides the technical grounding for the algorithms presented, and simulation results demonstrate how multiple agents using the proposed algorithms make a significant improvement in localization accuracy and speed over a single agent. Desirable landing sites for UAVs generally, and autonomous parafoils in particular, constitute locations where payload survivability and accessibility are greatest. An on-board semantic representation and understanding of the operating environment will allow systems to automatically detect and classify these desirable and undesirable landing sites, enabling successful landing in a much larger class of scenarios than currently possible, in addition to a fail-safe capability when localization is not feasible. This paper will further introduce an algorithm enabling multiple agents to reproduce this classification of their environment on-line better and more rapidly than a single agent working alone, demonstrating the advantage of multiple agents in this context as well. This has applications to general UAV landing, whether powered or unpowered, on Earth or in extra-terrestrial environments.