Multi-camera visual-inertial odometry: algorithm design and performance evaluation
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
It is essential for micro aerial vehicles (MAVs) to be able to track their 3D
motion in real time using limited onboard sensing and computational resources.
To this end, the Multi-State Constraint Kalman Filter (MSCKF), a variant of the
Extended Kalman Filter (EKF), is among the most popular approaches to provide
real-time motion tracking using camera and IMU sensors and has been proven capable of
accurate localization at reduced computational cost. As cameras are becoming
ubiquitous, one of the main objectives of this work is to design an efficient
multi-camera MSCKF-based visual-inertial odometry (VIO) algorithm that can
utilize an arbitrary number of asynchronous cameras. While the standard MSCKF
framework is well defined, it leaves available many design choices in
implementation that have significant impact on both localization accuracy and
computational complexity. These design choices depend strongly on the
computational resources available on the platform of interest, which can vary
greatly from platforms such as laptop computers to more resource-constrained
system-on-chip (SoC) platforms typically found onboard MAVs. We evaluate the
performance of our MSCKF-based VIO algorithm on a set of different computing
platforms in order to better understand how to optimize the estimator parameters
in a resource-aware way. These parameters include the number of stochastic
clones to retain in the state vector and the number of visual feature points to
track. In particular, in the design of our multi-camera VIO algorithm, the
number of independent camera streams to process is also considered as a design
parameter. The results of this study provide insight on selecting the parameters
that yield the best possible performance for a given computing platform.
Description
Keywords
Autonomous navigation, Kalman filter, Localization and mapping, State estimation, Unmanned aerial vehicle, Visual inertial odometry