Event-based real-time detection and tracking of UAVs

High-performance visual detection and tracking of small unmanned aerial vehicles (UAVs) at a distance is crucial for laser-based countermeasures, which require precise and accurate aiming to be effective. Achieving robust and accurate real-time tracking while dealing with dynamic non-cooperative targets and challenging lighting conditions poses significant difficulties for tracking solutions. Event cameras represent a new perception paradigm to overcome these hurdles by registering changes in illumination asynchronously with low latency, high temporal resolution and large dynamic range. The cameras yield a sparse stream of event information which encodes changes in the scene. Accumulating events into frames and using traditional computer vision techniques can be applied, but sacrifice some of the information contained in the event stream. Asynchronous methods have shown promise in utilizing the event stream more efficiently, but are not yet at the same technical maturity as frame-based alternatives. This work introduces and evaluates two event-based end-to-end detection and tracking systems suitable for real-time applications. One asynchronous and one accumulating method were studied. The different methods are evaluated using an annotated event-dataset containing recordings of multi-rotor UAVs operating in an outdoor setting to compare the performance.