Initialisation and Estimation Methods for Batch Optimisation of Inertial/Visual SLAM
(English)Manuscript (preprint) (Other academic)
Inertial/visual SLAM aims at estimating the pose of a camera and a map of landmarks detected in the images, using support from an inertial measurement unit (IMU). Some of the most competitive approaches for SLAM and its computer vision counterpart, Structure From Motion (SFM), are based on batch formulations such as GraphSLAM or Bundle Adjustment (BA). A major challenge in the implementation is the initialisation since these problems are inherently nonlinear and nonconvex. We propose a multi-step algorithm that solves a series of simple and almost uncoupled problems, often leading to linear solutions. It is believed that this leads to a robust algorithm which is simple to implement and fast to execute. The initialisation method is demonstrated on simulated data and a small feasibility study on experimental data using an industrial robot, to get access to ground truth, is also performed.
SLAM, Computer Vision, Inertial Sensors, Sensor Fusion, Visual Tracking
IdentifiersURN: urn:nbn:se:liu:diva-110370OAI: oai:DiVA.org:liu-110370DiVA: diva2:744964