TU Darmstadt / ULB / TUbiblio

Incorporating the Unscented Transform in Rao-Blackwellised Visual-Inertial SLAM

Borrmann, Daniel (2016):
Incorporating the Unscented Transform in Rao-Blackwellised Visual-Inertial SLAM.
Darmstadt, TU, Master Thesis, 2016, [Master Thesis]

Abstract

Monocular visual-inertial simultaneous localisation and mapping (SLAM) systems use a single camera and inertial sensors to build a sparse map of an environment and to simultaneously estimate the camera's position. Applications can be found, for example, in the robotics domain or in virtual and augmented reality applications. Nowadays, monocular visual SLAM systems are capable of operating in real-time on different platforms. The application of Rao-Blackwellised filtering techniques in visual SLAM has further improved the real-time capability and maximum map size of these systems. In many implementations, extended Kalman filters (EKFs) are used to estimate the camera's pose and landmark positions. Since the required dynamic and measurement functions in a visual SLAM system are usually non-linear, EKF-based implementations need to linearise these functions, which leads to linearisation errors in the state estimate. In these cases, the unscented Kalman filter (UKF) proved to provide better results in many scenarios. In this thesis, the unscented transform (UT) is incorporated into a Rao-Blackwellised visual-inertial SLAM system. It shows how the camera state and landmark positions can be estimated in unscented Rao-Blackwellised filtering. To this end, an appropriate motion model, as well as an inverse depth parametrisation for the landmark estimation, are incorporated into the filter. The proposed filter is tested on a visual-inertial dataset in different setups. Results are given for the UKF-based Rao- Blackwellised filter and an equivalent EKF-based implementation, to enable comparison between the two approaches.

Item Type: Master Thesis
Erschienen: 2016
Creators: Borrmann, Daniel
Title: Incorporating the Unscented Transform in Rao-Blackwellised Visual-Inertial SLAM
Language: English
Abstract:

Monocular visual-inertial simultaneous localisation and mapping (SLAM) systems use a single camera and inertial sensors to build a sparse map of an environment and to simultaneously estimate the camera's position. Applications can be found, for example, in the robotics domain or in virtual and augmented reality applications. Nowadays, monocular visual SLAM systems are capable of operating in real-time on different platforms. The application of Rao-Blackwellised filtering techniques in visual SLAM has further improved the real-time capability and maximum map size of these systems. In many implementations, extended Kalman filters (EKFs) are used to estimate the camera's pose and landmark positions. Since the required dynamic and measurement functions in a visual SLAM system are usually non-linear, EKF-based implementations need to linearise these functions, which leads to linearisation errors in the state estimate. In these cases, the unscented Kalman filter (UKF) proved to provide better results in many scenarios. In this thesis, the unscented transform (UT) is incorporated into a Rao-Blackwellised visual-inertial SLAM system. It shows how the camera state and landmark positions can be estimated in unscented Rao-Blackwellised filtering. To this end, an appropriate motion model, as well as an inverse depth parametrisation for the landmark estimation, are incorporated into the filter. The proposed filter is tested on a visual-inertial dataset in different setups. Results are given for the UKF-based Rao- Blackwellised filter and an equivalent EKF-based implementation, to enable comparison between the two approaches.

Uncontrolled Keywords: Guiding Theme: Digitized Work, Research Area: Computer vision (CV), Camera tracking, Inertial sensors, Filtering, Kalman filters
Divisions: 20 Department of Computer Science
20 Department of Computer Science > Mathematical and Applied Visual Computing
Date Deposited: 09 May 2019 10:33
Export:
Suche nach Titel in: TUfind oder in Google

Optionen (nur für Redakteure)

View Item View Item