Asynchronous Event-Based Visual Odometry for Space Exploration Robots

Abstract

Space exploration robots critically rely on their ability to estimate their motion and perceive their surroundings. While this presents a challenging problem for traditional planetary rovers, it becomes significantly more demanding for aerial platforms operating in environments like Mars. Navigating challenging extraterrestrial terrains, such as lava tubes and rocky landscapes, requires robust performance across a high dynamic range of lighting conditions. Simultaneously, these platforms are bound by strict mass, power, and onboard computational constraints. To address these challenges, this thesis presents a novel approach to visual odometry, using event cameras, leveraging the unique advantages these sensors offer for small, autonomous robots operating in such hostile scenarios. Unlike conventional frame-based methods, event cameras provide asynchronous, sparse, and high temporal resolution measurements. This enables robust perception even during high-speed maneuvers under extreme lighting conditions at low bandwidth. The proposed method utilizes tracked event features to continuously estimate camera motion at microsecond resolution. By registering events with a monocular event camera, the proposed pipeline computes odometry asynchronously following the Error State formulation for the Extended Kalman filter. Two different formulations are presented, which rely on the same architecture, and same front-end, but differ on how they initialize and parametrize the landmarks, seeking for robustness and real-time performance.