Extended Kalman Filter versus Newton-Lowe's Method for Robot Pose Estimation

摘要

In this work, we study the pose estimation problem of anautonomous mobile robot. Particularly, we compare the Extended Kalman Filter (EKF) to Lowe's method based on the iterative Newton's method for solving a system of nonlinear equations. Although the EKF is recursive which renders it suitable for the real-time problem at hand, Lowe's method has much less dimensionality. This is the motivation for comparing both approaches. We have used the stereo information for obtaining the 3-D structure and outlier rejection. This has provided an opportunity to weigh feeding both algorithms with single measurements (from one camera) against feeding them with pair measurements (from the stereo pair). We have studied the effects of using three ranges of the number of features and the longevity on the accuracy of the obtained pose parameters. Moreover, we have investigated the impact of the number of iterations on the accuracy of Lowe's method. An extensive set of simulations as well as real experiments using various motion patterns have been conducted. The main finding of this work is that Lowe's method (due to its low dimensionality) is much faster with approximately the same accuracy. Besides, it can recover from a situation which is close to singularity. On the other hand, the EKF makes better use of multiple camera measurements which allows a sustained performance even if one camera is off or occluded.