## Square Root SAM: Simultaneous localization and mapping via square root information smoothing (2006)

### Cached

### Download Links

Venue: | International Journal of Robotics Reasearch |

Citations: | 91 - 32 self |

### BibTeX

@ARTICLE{Dellaert06squareroot,

author = {Frank Dellaert and Michael Kaess},

title = {Square Root SAM: Simultaneous localization and mapping via square root information smoothing},

journal = {International Journal of Robotics Reasearch},

year = {2006},

volume = {25},

pages = {2006}

}

### Years of Citing Articles

### OpenURL

### Abstract

Solving the SLAM problem is one way to enable a robot to explore, map, and navigate in a previously unknown environment. We investigate smoothing approaches as a viable alternative to extended Kalman filter-based solutions to the problem. In particular, we look at approaches that factorize either the associated information matrix or the measurement Jacobian into square root form. Such techniques have several significant advantages over the EKF: they are faster yet exact, they can be used in either batch or incremental mode, are better equipped to deal with non-linear process and measurement models, and yield the entire robot trajectory, at lower cost for a large class of SLAM problems. In addition, in an indirect but dramatic way, column ordering heuristics automatically exploit the locality inherent in the geographic nature of the SLAM problem. In this paper we present the theory underlying these methods, along with an interpretation of factorization in terms of the graphical model associated with the SLAM problem. We present both simulation results and actual SLAM experiments in large-scale environments that underscore the potential of these methods as an alternative to EKF-based approaches. 1

### Citations

3374 |
Multiple View Geometry in Computer Vision
- Hartley, Zisserman
- 2000
(Show Context)
Citation Context ...arameters of all features in the environment. In fact, this problem has a long history in surveying [36], photogrammetry [7, 37, 68, 10], where it is known as “bundle adjustment”, and computer vision =-=[25, 72, 73, 79, 41]-=-, where it is referred to as “structure from motion”. Especially in the last five years there has been a flurry of work where these ideas were applied in the context of SLAM [16, 22, 23, 43, 30, 31, 2... |

1308 |
Three-Dimensional Computer Vision: A Geometric Viewpoint
- Faugeras
- 1993
(Show Context)
Citation Context ...arameters of all features in the environment. In fact, this problem has a long history in surveying [36], photogrammetry [7, 37, 68, 10], where it is known as “bundle adjustment”, and computer vision =-=[25, 72, 73, 79, 41]-=-, where it is referred to as “structure from motion”. Especially in the last five years there has been a flurry of work where these ideas were applied in the context of SLAM [16, 22, 23, 43, 30, 31, 2... |

1260 | Factor graphs and the sum-product algorithm
- Kschischang, Frey, et al.
- 2001
(Show Context)
Citation Context ...e graphical models that can be used to represent such factored (2) (3)s2 SLAM AND ITS GRAPHS 6 Figure 3: The undirected Markov random fields associated with the SLAM problems from Figure 1. densities =-=[48]-=-. In a factor graph there are nodes for unknowns and nodes for the probability factors defined on them, and the graph structure expresses which unknowns are involved in each factor. The factor graphs ... |

987 |
Numerical methods for unconstrained optimization and nonlinear equations
- Dennis, Schnabel
- 1996
(Show Context)
Citation Context ...vailable, non-linear optimization methods such as Gauss-Newton iterations or the Levenberg-Marquardt algorithm will solve a succession of linear approximations to (5) in order to approach the minimum =-=[20]-=-. This is similar to the extended Kalman filter approach to SLAM as pioneered by [70, 71, 52], but allows for iterating multiple times to convergence while controlling in which region one is willing t... |

854 |
Probabilistic Robotics
- Thrun, Burgard, et al.
- 2005
(Show Context)
Citation Context ...onment itself is the artifact of interest, e.g. in urban reconstruction, search-and-rescue operations, battlefield reconnaissance etc. As such, it is one of the core competencies of autonomous robots =-=[77]-=-. We will primarily be concerned with landmark-based SLAM, for which the earliest and most popular methods are based on the extended Kalman filter (EKF) [70, 62, 61, 2, 71, 53, 51]. The EKF recursivel... |

654 |
Probabilistic Networks and Expert Systems
- Cowell, Dawid, et al.
- 1999
(Show Context)
Citation Context ...state of the art in linear algebra is a blend of numerical methods and advanced graph theory, sharing many characteristics with inference algorithms typically found in the graphical models literature =-=[11]-=-. In particular, the graph-theoretic algorithm underlying the solution of sparse least-squares is variable elimination, in which each variable (such as a robot pose or a landmark position) is expresse... |

470 | FastSLAM: A factored solution to the simultaneous localization and mapping problem
- Montemerlo, Thrun, et al.
- 2002
(Show Context)
Citation Context ...m when ordering the variables. Looking at SLAM in terms of graphs has a rich history in itself [6, 74, 33, 34] and has especially lead to several novel and exciting developments in the past few years =-=[63, 60, 30, 65, 26, 31, 78, 27, 77]-=-. Below we more closely examine the tight connection between the graphical model view and the sparse linear algebra formulation of SLAM. It is well known that the information matrix I is associated wi... |

452 | Globally consistent range scan alignment for environment mapping. Autonomous Robots
- Lu, Milios
- 1997
(Show Context)
Citation Context ...ch to SLAM involves not just the most current robot location, but the entire robot trajectory up to the current time. A number of authors considered the problem of smoothing the robot trajectory only =-=[9, 56, 57, 40, 46, 24]-=-, which is particularly suited to sensors such as laserrange finders that easily yield pairwise constraints between nearby robot poses. More generally, one can consider the full SLAM problem [77], i.e... |

415 | Bundle adjustment – a modern synthesis
- Triggs, McLauchlan, et al.
- 2000
(Show Context)
Citation Context ...arameters of all features in the environment. In fact, this problem has a long history in surveying [36], photogrammetry [7, 37, 68, 10], where it is known as “bundle adjustment”, and computer vision =-=[25, 72, 73, 79, 41]-=-, where it is referred to as “structure from motion”. Especially in the last five years there has been a flurry of work where these ideas were applied in the context of SLAM [16, 22, 23, 43, 30, 31, 2... |

409 | Generalized belief propagation
- Yedidia, Freeman, et al.
- 2000
(Show Context)
Citation Context ... by a common factor (measurement or constraint). At this level of abstraction, the form (4)s3 SAM AS A LEAST SQUARES PROBLEM 7 corresponds exactly to the expression for a pairwise Markov random field =-=[82]-=-, hence MRFs and factor graphs are equivalent representations here. The MRFs for the examples from Figure 1 are shown in Figure 4. Note that it looks very similar to Figure 2, but an MRF is a fundamen... |

408 |
Estimating uncertain spatial relationships in Robotics
- Smith, Self, et al.
- 1990
(Show Context)
Citation Context ...the core competencies of autonomous robots [77]. We will primarily be concerned with landmark-based SLAM, for which the earliest and most popular methods are based on the extended Kalman filter (EKF) =-=[70, 62, 61, 2, 71, 53, 51]-=-. The EKF recursively estimates a Gaussian density over the current pose of the robot and the position of all landmarks (the map). However, it is well known that the computational complexity of the EK... |

403 | A separator theorem for planar graphs
- Lipton, Tarjan
- 1979
(Show Context)
Citation Context ...medium-sized problems is colamd [1], which works on the columns of A. Another popular method, based on graph theory and often used to speed up finite element methods, is generalized nested dissection =-=[55, 54]-=-. Given that an optimal ordering is out of reach in general, heuristics or domain-knowledge can do much better than general-purpose algorithms. A simple idea is to use a standard method such as colamd... |

376 | A solution to simultaneous localization and map building (slam) problem
- Clark, Dissanayake, et al.
- 2001
(Show Context)
Citation Context ...he EKF becomes intractable fairly quickly, and hence a large number of efforts have focused on 1s1 INTRODUCTION 2 modifying and extending the filtering approach to cope with larger-scale environments =-=[64, 17, 21, 8, 45, 49, 38, 65, 50, 5, 78, 39, 75, 80, 67]-=-. However, filtering itself has been shown to be inconsistent when applied to the inherently non-linear SLAM problem [44], i.e., even the average taken over a large number of experiments diverges from... |

313 |
On the representation and estimation of spatial uncertainty
- Smith, Cheeseman
- 1986
(Show Context)
Citation Context ...iments in large-scale environments that underscore the potential of these methods as an alternative to EKF-based approaches. 1 Introduction The problem of simultaneous localization and mapping (SLAM) =-=[69, 51, 76]-=- has received considerable attention in mobile robotics as it is one way to enable a robot to explore and navigate previously unknown environments. In addition, in many applications the map of the env... |

294 | Incremental mapping of large cyclic environments
- Gutmann, Konolige
- 1999
(Show Context)
Citation Context ...ch to SLAM involves not just the most current robot location, but the entire robot trajectory up to the current time. A number of authors considered the problem of smoothing the robot trajectory only =-=[9, 56, 57, 40, 46, 24]-=-, which is particularly suited to sensors such as laserrange finders that easily yield pairwise constraints between nearby robot poses. More generally, one can consider the full SLAM problem [77], i.e... |

265 | A column approximate minimum degree ordering algorithm
- Davis, Gilbert, et al.
- 2004
(Show Context)
Citation Context ...uch as QR and Cholesky factorization. Finding an optimal ordering is an NP-complete problem, but there are several ordering heuristics and approximate algorithms that perform well on general problems =-=[1, 42]-=- and are built into programs like MATLAB [58]. While general-purpose ordering methods drastically improve performance, yet another orderof-magnitude improvement can be obtained by exploiting the speci... |

257 | Directed Sonar Sensing for Mobile Robot Navigation
- Leonard, Durrant-Whyte
- 1992
(Show Context)
Citation Context ...the core competencies of autonomous robots [77]. We will primarily be concerned with landmark-based SLAM, for which the earliest and most popular methods are based on the extended Kalman filter (EKF) =-=[70, 62, 61, 2, 71, 53, 51]-=-. The EKF recursively estimates a Gaussian density over the current pose of the robot and the position of all landmarks (the map). However, it is well known that the computational complexity of the EK... |

252 |
Stochastic Models, Estimation, and Control, volume 141-2 ofMathematics in Science and Engineering
- Maybeck
- 1982
(Show Context)
Citation Context ...n JPL’s Mariner 10 missions to Venus (as recounted by [3]). The use of square roots of either the covariance or information matrix results in more accurate and stable algorithms, and, quoting Maybeck =-=[59]-=- “a number ofs2 SLAM AND ITS GRAPHS 3 practitioners have argued, with considerable logic, that square root filters should always be adopted in preference to the standard Kalman filter recursion”. Mayb... |

241 | Robot pose estimation in unknown environments by matching 2D range scans
- Lu, Milios
- 1997
(Show Context)
Citation Context ...ch to SLAM involves not just the most current robot location, but the entire robot trajectory up to the current time. A number of authors considered the problem of smoothing the robot trajectory only =-=[9, 56, 57, 40, 46, 24]-=-, which is particularly suited to sensors such as laserrange finders that easily yield pairwise constraints between nearby robot poses. More generally, one can consider the full SLAM problem [77], i.e... |

222 |
Image Analysis, Random Fields, and Dynamic Monte Carlo Methods
- Winkler
- 1995
(Show Context)
Citation Context ...Again, at the block-level the sparsity pattern of A T A is exactly the adjacency matrix of the associated MRF. The objective function in Equation 5 corresponds to a pairwise Markov random field (MRF) =-=[81, 82]-=- through the Hammersley-Clifford theorem [81], and the nodes in the MRF correspond to the robot states and the landmarks. Links represent either odometry or landmark measurements. In [65, 78] the MRF ... |

200 |
Simultaneous localization and mapping with sparse extended information filters
- Thrun, Liu, et al.
- 2004
(Show Context)
Citation Context ...graph that is examined by general purpose linear algebra methods. By working with the block-structure instead, we will be able to do better. (12)s5 A GRAPHICAL MODEL PERSPECTIVE 12 As noted before in =-=[78]-=- and by others, the information matrix I = A T A is the matrix of the Markov random field associated with the SLAM problem. Again, at the block-level the sparsity pattern of A T A is exactly the adjac... |

197 | Recovering 3D Shape and Motion From Image Streams Using Nonlinear Least Squares
- Szeliski
- 1994
(Show Context)
Citation Context |

190 |
Generalized nested dissection
- Lipton, Tarjan
- 1979
(Show Context)
Citation Context ...medium-sized problems is colamd [1], which works on the columns of A. Another popular method, based on graph theory and often used to speed up finite element methods, is generalized nested dissection =-=[55, 54]-=-. Given that an optimal ordering is out of reach in general, heuristics or domain-knowledge can do much better than general-purpose algorithms. A simple idea is to use a standard method such as colamd... |

189 |
Dynamic map building for autonomous mobile robot
- Leonard, Durrant-Whyte, et al.
- 1990
(Show Context)
Citation Context ...iments in large-scale environments that underscore the potential of these methods as an alternative to EKF-based approaches. 1 Introduction The problem of simultaneous localization and mapping (SLAM) =-=[69, 51, 76]-=- has received considerable attention in mobile robotics as it is one way to enable a robot to explore and navigate previously unknown environments. In addition, in many applications the map of the env... |

188 | Optimization of simultaneous localization and map building algorithm for real time implementation
- Guivant, Nebot
- 2001
(Show Context)
Citation Context ...he EKF becomes intractable fairly quickly, and hence a large number of efforts have focused on 1s1 INTRODUCTION 2 modifying and extending the filtering approach to cope with larger-scale environments =-=[64, 17, 21, 8, 45, 49, 38, 65, 50, 5, 78, 39, 75, 80, 67]-=-. However, filtering itself has been shown to be inconsistent when applied to the inherently non-linear SLAM problem [44], i.e., even the average taken over a large number of experiments diverges from... |

187 |
Position referencing and consistent world modeling for mobile robots
- Chatila, Laumond
- 1985
(Show Context)
Citation Context |

162 |
Simultaneous map building and localization for an autonomous mobile robot
- Leonard, Durrant-Whyte
- 1991
(Show Context)
Citation Context ...nberg-Marquardt algorithm will solve a succession of linear approximations to (5) in order to approach the minimum [20]. This is similar to the extended Kalman filter approach to SLAM as pioneered by =-=[70, 71, 52]-=-, but allows for iterating multiple times to convergence while controlling in which region one is willing to trust the linear assumption (hence, these methods are often called region-trust methods). W... |

154 |
Maintaining representations of the environment of a mobile robot
- Ayache, Faugeras
- 1989
(Show Context)
Citation Context ...the core competencies of autonomous robots [77]. We will primarily be concerned with landmark-based SLAM, for which the earliest and most popular methods are based on the extended Kalman filter (EKF) =-=[70, 62, 61, 2, 71, 53, 51]-=-. The EKF recursively estimates a Gaussian density over the current pose of the robot and the position of all landmarks (the map). However, it is well known that the computational complexity of the EK... |

150 | An Atlas framework for scalable mapping
- Bosse, Newman, et al.
- 2003
(Show Context)
Citation Context ...he EKF becomes intractable fairly quickly, and hence a large number of efforts have focused on 1s1 INTRODUCTION 2 modifying and extending the filtering approach to cope with larger-scale environments =-=[64, 17, 21, 8, 45, 49, 38, 65, 50, 5, 78, 39, 75, 80, 67]-=-. However, filtering itself has been shown to be inconsistent when applied to the inherently non-linear SLAM problem [44], i.e., even the average taken over a large number of experiments diverges from... |

146 | Robust Mapping and Localization in Indoor Environments using Sonar Data
- Tardos, Neira, et al.
- 2002
(Show Context)
Citation Context |

143 | Bayesian map learning in dynamic environments
- Murphy
- 2000
(Show Context)
Citation Context ...m when ordering the variables. Looking at SLAM in terms of graphs has a rich history in itself [6, 74, 33, 34] and has especially lead to several novel and exciting developments in the past few years =-=[63, 60, 30, 65, 26, 31, 78, 27, 77]-=-. Below we more closely examine the tight connection between the graphical model view and the sparse linear algebra formulation of SLAM. It is well known that the information matrix I is associated wi... |

142 |
Factorization Methods for Discrete Sequential Estimation
- Bierman
- 1977
(Show Context)
Citation Context ...formation matrix is known in the sequential estimation literature as square root information filtering (SRIF), and was developed in 1969 for use in JPL’s Mariner 10 missions to Venus (as recounted by =-=[3]-=-). The use of square roots of either the covariance or information matrix results in more accurate and stable algorithms, and, quoting Maybeck [59] “a number ofs2 SLAM AND ITS GRAPHS 3 practitioners h... |

140 | A stochastic map for uncertain spatial relationships
- Smith, Self, et al.
- 1988
(Show Context)
Citation Context |

129 | Tardós, “The SPmap: a probabilistic framework for simultaneous localization and map building
- Castellanos, Montiel, et al.
- 1999
(Show Context)
Citation Context |

126 | Thin junction tree filters for simultaneous localization and mapping
- Paskin
(Show Context)
Citation Context |

115 |
An introduction to chordal graphs an clique trees, The IMA volumes
- Blair, Peyton
- 1993
(Show Context)
Citation Context ...a junction tree, known from inference in graphical models and also recently applied in SLAM [65]. Both factorization methods, QR and Cholesky (or LDL), are based on the variable elimination algorithm =-=[4, 11]-=-. The difference between these methods is that QR eliminates variable nodes from the factor graph and obtains A = QR, while Cholesky or LDL start from the MRF and hence obtains5 A GRAPHICAL MODEL PERS... |

92 | A multilevel relaxation algorithm for simultaneous localization and mapping
- Frese, Larsson, et al.
- 2005
(Show Context)
Citation Context ...ision [25, 72, 73, 79, 41], where it is referred to as “structure from motion”. Especially in the last five years there has been a flurry of work where these ideas were applied in the context of SLAM =-=[16, 22, 23, 43, 30, 31, 26, 27, 28, 29, 77]-=-. In this paper we show how smoothing can be a very fast alternative to filtering-based methods, and that in many cases keeping the trajectory around helps rather than hurts performance. In particular... |

79 | Experimental comparison of techniques for localization and mapping using a bearings only sensor
- Deans, Hebert
(Show Context)
Citation Context ...ision [25, 72, 73, 79, 41], where it is referred to as “structure from motion”. Especially in the last five years there has been a flurry of work where these ideas were applied in the context of SLAM =-=[16, 22, 23, 43, 30, 31, 26, 27, 28, 29, 77]-=-. In this paper we show how smoothing can be a very fast alternative to filtering-based methods, and that in many cases keeping the trajectory around helps rather than hurts performance. In particular... |

79 | A counter example to the theory of simultaneous localization and map building
- Julier, Uhlmann
(Show Context)
Citation Context ...arger-scale environments [64, 17, 21, 8, 45, 49, 38, 65, 50, 5, 78, 39, 75, 80, 67]. However, filtering itself has been shown to be inconsistent when applied to the inherently non-linear SLAM problem =-=[44]-=-, i.e., even the average taken over a large number of experiments diverges from the true solution. Since this is mainly due to linearization choices that cannot be undone in a filtering framework, the... |

72 | Relaxation on a mesh: a formalism for generalized localization
- Howard, Matarić, et al.
- 2001
(Show Context)
Citation Context ...ision [25, 72, 73, 79, 41], where it is referred to as “structure from motion”. Especially in the last five years there has been a flurry of work where these ideas were applied in the context of SLAM =-=[16, 22, 23, 43, 30, 31, 26, 27, 28, 29, 77]-=-. In this paper we show how smoothing can be a very fast alternative to filtering-based methods, and that in many cases keeping the trajectory around helps rather than hurts performance. In particular... |

68 |
Large-scale map-making
- Konolige
(Show Context)
Citation Context |

64 | Learning globally consistent maps by relaxation
- Duckett, Marsland, et al.
- 2000
(Show Context)
Citation Context |

62 | Exactly sparse delayed-state filters
- Eustice, Singh, et al.
- 2005
(Show Context)
Citation Context |

61 |
On the Structure and Solution of the Simultaneous Localisation and Map Building Problem
- Newman
- 1999
(Show Context)
Citation Context |

58 |
Visual map making for a mobile robot
- Brooks
- 1985
(Show Context)
Citation Context ...f-magnitude improvement can be obtained by exploiting the specific graphical structure of the SLAM problem when ordering the variables. Looking at SLAM in terms of graphs has a rich history in itself =-=[6, 74, 33, 34]-=- and has especially lead to several novel and exciting developments in the past few years [63, 60, 30, 65, 26, 31, 78, 27, 77]. Below we more closely examine the tight connection between the graphical... |

57 |
Robotic mapping: A survey,” in Exploring artificial intelligence in the new millennium
- Thrun
- 2003
(Show Context)
Citation Context ...iments in large-scale environments that underscore the potential of these methods as an alternative to EKF-based approaches. 1 Introduction The problem of simultaneous localization and mapping (SLAM) =-=[69, 51, 76]-=- has received considerable attention in mobile robotics as it is one way to enable a robot to explore and navigate previously unknown environments. In addition, in many applications the map of the env... |

56 |
An experimental system for incremental environment modelling by an autonomous mobile robot
- Moutarlier, Chatila
- 1989
(Show Context)
Citation Context |

54 | Graphical slam: A self-correcting map
- Folkesson, Christensen
- 2004
(Show Context)
Citation Context |

54 | Decoupled stochastic mapping
- Leonard, Feder
- 2001
(Show Context)
Citation Context |

51 |
Square Root SAM: Simultaneous location and mapping via square root information smoothing
- Dellaert
- 2005
(Show Context)
Citation Context ...blem of simultaneous smoothing and mapping (SAM). Because they are based on matrix square roots, we will refer to this family of approaches as square root SAM, or √ SAM for short, first introduced in =-=[18]-=-. We propose that √ SAM is a fundamentally better approach to the problem of SLAM than the EKF, based on the realization that, • in contrast to the extended Kalman filter covariance or information mat... |