Event-based Particle Filtering for Robot Self-Localization

We propose a novel algorithm for robot self-localization using an embedded event-based sensor. This sensor produces a stream of events at microsecond time resolution which only represents pixel-level illumination changes in a scene, as e. g. caused by perceived motion. This is in contrast to classical image sensors, which wastefully transmit redundant information at a much lower frame rate. Our method adapts the commonly used Condensation Particle Filter Tracker to such event-based sensors. It works directly with individual, highly ambiguous pixel-events and does not employ event integration over time. The lack of complete discrete sensory measurements is addressed by applying an exponential decay model for hypotheses likelihood computation. The proposed algorithm demonstrates robust performance at low computation requirements; turning it suitable for implementation in embedded hardware on small autonomous robots. We evaluate our algorithm in a simulation environment and with experimental recorded data.