Robust incremental SLAM with consistency-checking

Abstract

Abstract-Both landmark measurements and loop closures are used to correct for odometry drift in SLAM solutions. However, if any of the measurements are incorrect (e.g. due to perceptual aliasing) standard SLAM algorithms fail catastrophically and can not return an accurate map. A number of algorithms have been proposed that are robust to loop closure errors, but it is shown in this paper that they can not provide robust solutions when landmark measurement errors occur. The root cause of the problem is that most robust SLAM algorithms only focus on creating a locally consistent map (by evaluating whether measurements appear correct individually) rather than a globally consistent map. This paper proposes a new formulation of the robust SLAM problem that explicitly requires finding a globally consistent solution. Motivated by the new cost function, a novel incremental SLAM algorithm is developed that provides accurate solutions for datasets with landmark or loop closure measurement errors. Simulated and experimental results of the new algorithm, called incremental SLAM with consistency-checking, show that the new algorithm provides significantly more accurate results than state-of-theart robust SLAM methods for datasets with incorrect landmark measurements and can match the performance of current robust SLAM methods for datasets with incorrect loop closures.

A nonparametric belief solution to the Bayes tree

“…Recently in the SLAM community Graham [4] used an expectation maximization (EM) approach to smoothly transition poorly matched measurements to assumed outliers. The EM algorithm iterates between covariance weight selection and optimal variable assignments and suppresses "misbehaved" measurements by emphasizing the majority of constraints with consensus.…”

Convex Relaxations for Pose Graph Optimization With Outliers

“…Similar ideas appear in the context of robust estimation, e.g., the Penalized Trimmed Squares estimator of Zioutas and Avramidis, [33]. Latif et al in [20] and Graham et al in [11] look for "internally consistent" constraints, which are in mutual agreement. Carlone et al in [6] use 1 -relaxation to find a large set of mutually-consistent measurements.…”

Multi-modal and inertial sensor solutions for navigation-type factor graphs

“…More recently, Graham [76] suggests using an expectation maximization (EM) approach to smoothly transition poorly matched measurements to assumed "outliers" by adjusting their measurement covariance. The EM algorithm is used to iterate between covariance weight selection and optimal variable assignments and thereby suppresses outlier-like measurements and emphasizing the majority of constraints which form consensus.…”

“…For example, consider a point on the lower right just above the unimodal parametric curve: (i) What is the complexity of tracking a limited number of hypotheses across the entire solution (FastSLAM [81]) versus tracking more modes in regions of the solution? (ii) Can we do better than assumed "outlier" approaches where measurements are de-weighted as null-hypotheses, such as switch type variables [76,206,219]? (iii) Can we select a piece of the problem where clear uncertainties exist and then treat just that portion in some special manner, leaving the rest as a conventional max-product parametric solution?…”