Analysis of Different Feature Selection Criteria Based on a Covariance Convergence Perspective for a SLAM Algorithm

Abstract

This paper introduces several non-arbitrary feature selection techniques for a Simultaneous Localization and Mapping (SLAM) algorithm. The feature selection criteria are based on the determination of the most significant features from a SLAM convergence perspective. The SLAM algorithm implemented in this work is a sequential EKF (Extended Kalman filter) SLAM. The feature selection criteria are applied on the correction stage of the SLAM algorithm, restricting it to correct the SLAM algorithm with the most significant features. This restriction also causes a decrement in the processing time of the SLAM. Several experiments with a mobile robot are shown in this work. The experiments concern the map reconstruction and a comparison between the different proposed techniques performance. The experiments were carried out at an outdoor environment composed by trees, although the results shown herein are not restricted to a special type of features.

Covariance-Based Measurement Selection Criterion for Gaussian-Based Algorithms

“…Let z ∈ N be a feature -observation-from the environment and R the covariance matrix of the corresponding observation noise-obtained from the feature extraction procedure. Equation (3) implies that the most significant feature in N at time k is the one that causes the largest decrement of the uncertainty volume of the corrected covariance matrix of the Gaussian estimation algorithm (see [29]). Therefore, the selection criterion implemented in this work consists of using only the optimal features found in Equation (3) to correct the estimation algorithm.…”

“…Thus, under the same hypothesis of Lemma V, for the EKF case: |P | Ra ≥ |P | R b ; and, for the EIF case: |Ω| Ra ≤ |Ω| R b . Both expressions can be directly obtained from Equations (2.1) and (12), considering that a same feature has a same Jacobian matrix associated with it [29].…”

“…The off-line execution of the SLAM algorithm allows the study of the SLAM behavior without considering the motion implications of the mobile robot controller nor its inherent delays ( [29,38] use real time SLAM executions with non-reactive mobile robot motion controllers).…”

Real time simultaneous localization and mapping: towards low-cost multiprocessor embedded systems

“…The new state vector includes new landmarks (whose initialization has just been performed) and previously used landmarks. Auat Cheein and Carelli [26] proposes an efficient method to select landmarks for the estimation task. It is based on the evaluation of the influence of a given feature on the convergence of the state covariance matrix.…”

“…The method matches all possible landmarks and computes (I − K k H k ) from Equation (12). Unfortunately, we cannot implement it exactly as proposed by [26] due to the high computing time. We chose to add the landmarks, based on the previous estimation step, by selecting the previous landmarks which have the best previous influence on the convergence of the state covariance matrix.…”