Multi-Modal Landmark Detection and Tracking for Odometry Estimation in Degraded Visual Environments
StatisticsView Usage Statistics
This thesis focuses on the development of a multi-modal odometry estimation framework for the purpose of robot localization and mapping in visually degraded environments. Visual and depth information are encoded at the feature detection and descriptor extraction level, providing robustness to the landmark selection process in low illumination and texture-less conditions. An extended Kalman filter framework is used to predict landmark positions using inertial measurements in successive frames and the matching pixel error is used as an innovation term. For localization performance evaluation the proposed approach is compared to ground-truth provided by Vicon system and a state-of-the-art visual-inertial odometry estimation framework. The mapping performance is demonstrated by mapping a large room in dark conditions.