Multisensor SLAM for Indoor Navigation

Abstract

Affordable RGB-D sensors, such as Microsoft Kinect or ASUS Xtion Pro Live, have become a popular choice for gesture recognition and 3D reconstruction. However, in applications for Simultaneous Localization and Mapping (SLAM) the limited range and field of view of these sensors poses a challenge with respect to achieving accurate visual odometry. This thesis presents a multisensor setup for the Rollin’ Justin service robot platform and an approach to robust 3D mapping that is suitable for indoor navigation. Via calibration of multiple sensors the measurements can be projected into a common reference coordinate system for further processing. Matching the combined data provides better motion estimates than establishing the robot pose based on the measurements of the individual sensors. Furthermore, different algorithms for RGB-D frame alignment and techniques for minimizing the accumulated drift are studied. A tree-based map structure and signed distance fields allows one to integrate noisy measurements efficiently and store large areas with a low memory footprint. Moreover, the drift of frame-to-frame registration can be minimized through pose refinement by tracking the sensor frame against the map representation. The approach is validated by comparison to state-of-the-art methods on publicly available ground truth data and experiments on the Rollin’ Justin service robot.