Three Flavors of RGB-D Visual Odometry: Analysis of Cost Function Compromises and Covariance Estimation Accuracy

Prashant Ganesh, Kyle Volle, Andrew R. Willis, Kevin M. Brink

Abstract: Bayesian estimation hinges on characterization of measurement uncertainty. In this work we investigate classical and recent approaches for characterizing uncertainty in pose change estimation for several visual odometry (VO) algorithms, which can then be used in autonomous navigation or other estimation problems. Visual odometry algorithms estimate the relative 3D motion, i.e., pose-change, based on the change in scenery captured by a time sequence of color and, more recently, depth images. We compare three different algorithms (two featurebased and one pixel-based) on the basis of accuracy, robustness and reported uncertainty. Each method adopts a semanticallydistinct yet physically-meaningful cost function to solve for posechange and this work characterizes the trade-offs associated with the respective cost function formulation and algorithmic solution. We then discuss how to characterize the uncertainty of their resulting estimates using both classical theory, e.g., Fisher Information matrix, and statistical approaches, e.g., the bootstrap technique. Results highlight the ability of these approaches to create both an accurate measurement and, critically, a relevant per-measurement covariance value appropriate for use in Bayesian estimators with applications in guidance, navigation and control (GNC).
Published in: 2020 IEEE/ION Position, Location and Navigation Symposium (PLANS)
April 20 - 23, 2020
Hilton Portland Downtown
Portland, Oregon
Pages: 1587 - 1595
Cite this article: Ganesh, Prashant, Volle, Kyle, Willis, Andrew R., Brink, Kevin M., "Three Flavors of RGB-D Visual Odometry: Analysis of Cost Function Compromises and Covariance Estimation Accuracy," 2020 IEEE/ION Position, Location and Navigation Symposium (PLANS), Portland, Oregon, April 2020, pp. 1587-1595.
Full Paper: ION Members/Non-Members: 1 Download Credit
Sign In