## Scalable monocular SLAM (2006)

Venue: | in IEEE Computer Society Conference on Computer Vision and Pattern Recognition |

Citations: | 87 - 3 self |

### BibTeX

@INPROCEEDINGS{Eade06scalablemonocular,

author = {Ethan Eade and Tom Drummond},

title = {Scalable monocular SLAM},

booktitle = {in IEEE Computer Society Conference on Computer Vision and Pattern Recognition},

year = {2006},

pages = {469--476},

publisher = {IEEE Computer Society}

}

### Years of Citing Articles

### OpenURL

### Abstract

Localization and mapping in unknown environments becomes more difficult as the complexity of the environment increases. With conventional techniques, the cost of maintaining estimates rises rapidly with the number of landmarks mapped. We present a monocular SLAM system that employs a particle filter and top-down search to allow realtime performance while mapping large numbers of landmarks. To our knowledge, we are the first to apply this FastSLAM-type particle filter to single-camera SLAM. We also introduce a novel partial initialization procedure that efficiently determines the depth of new landmarks. Moreover, we use information available in observations of new landmarks to improve camera pose estimates. Results show the system operating in real-time on a standard workstation while mapping hundreds of landmarks. 1.

### Citations

1147 | A tutorial on particle filters for online nonlinear/non-gaussian bayesian tracking
- Arulampalam, Maskell, et al.
(Show Context)
Citation Context ...which observations are processed does not affect the cumulative result. After all observations are processed, poses are randomly sampled from the gaussian mixture. Using standard resampling techniques=-=[1]-=-, we create zero or more descendants from each particle according to its weight, with each descendant’s pose sampled from the particle’s associated gaussian. Data copying is minimized in the resamplin... |

426 | New Extension of the Kalman Filter to Nonlinear Systems
- Julier, Uhlmann
- 1997
(Show Context)
Citation Context ...ion of the corresponding x∗ is gaussian, the landmark is added to the particle’s map as a fully initialized three-dimensional point. The change of variables is performed using the Unscented Transform =-=[19]-=-, avoiding systematic bias that simply transforming the mean would induce. One additional concern in partial initialization is that depth estimates of new landmarks converge artificiallywhen no new i... |

256 | Recursive estimation of motion structure and focal length
- Azarbayejani, Pentland
- 1995
(Show Context)
Citation Context ...y the EKF. In the SFM literature, there has been significant success using the EKF for causal estimation – estimation depending only on observations up to the current time – with recursive algorithms =-=[2, 4, 9]-=-. In contrast to SFM approaches that rely on global nonlinear optimization, recursive estimation methods permit online operation, which is highly desirable for a SLAM system. Davison shows the feasibi... |

186 |
Simultaneous localization and mapping with sparse extended information filters,” Intl
- Thrun, Liu, et al.
- 2004
(Show Context)
Citation Context ... a method exists for inferring migration from one submap to another. Other approaches enforce sparsity of correlation between landmarks in an adaptive manner, either by sparsifying inverse covariance =-=[21]-=- or by choosing a different representation of the covariance that is adaptively compressed [16]. More recently, Montemerlo et. al [14] have exploited a probabilistic property inherent to the SLAM prob... |

183 | Optimization of the simultaneous localization and map building algorithm for real time implementation, in
- Guivant, Nebot
- 2001
(Show Context)
Citation Context ...time operation is desired. This level of map complexity allows localization and sparse mapping in a single room, but is not suited to very large areas or densely populated maps. Aggregated EKF updates=-=[8, 10]-=- allow efficient operation while observing a working set of landmarks, but full O(N 2 ) updates are still required when changing the working set. Since the number of landmarks N grows with time, O(N 2... |

165 | Fastslam 2.0: An Improved Particle Filtering Algorithm for Simultaneous Localization and Mapping that Provably Converges
- Montemerlo, Thrun, et al.
- 2003
(Show Context)
Citation Context ...as similar as possible to the desired posterior. In order to satisfy this criterion, proposal distributions should take into account the latest measurements available to the system [22]. FastSLAM 2.0 =-=[15]-=- shows how such improved proposals can be generated within the FastSLAM framework by adjusting the sampled poses according to new measurements. In order to successfully operate with few particles, thi... |

148 | An ATLAS framework for scalable mapping
- Bosse, Newman, et al.
- 2003
(Show Context)
Citation Context ...ally initialized landmarks to help constrain camera pose. 2. Background 2.1. Scalable SLAM Extensive work has been undertaken in the robotics community to address the complexity of large scale mapping=-=[3, 7, 13, 23]-=-. Several approaches explicitly model the weak covariance between geographically distant landmarks by fully decorrelating their estimates in submaps. In a submap of bounded complexity, computation and... |

145 | The unscented particle filter
- Merwe, Doucet, et al.
- 2000
(Show Context)
Citation Context ...sense that they be as similar as possible to the desired posterior. In order to satisfy this criterion, proposal distributions should take into account the latest measurements available to the system =-=[22]-=-. FastSLAM 2.0 [15] shows how such improved proposals can be generated within the FastSLAM framework by adjusting the sampled poses according to new measurements. In order to successfully operate with... |

122 | Thin Junction Tree Filters for Simultaneous Localization and
- Paskin
- 2003
(Show Context)
Citation Context ...parsity of correlation between landmarks in an adaptive manner, either by sparsifying inverse covariance [21] or by choosing a different representation of the covariance that is adaptively compressed =-=[16]-=-. More recently, Montemerlo et. al [14] have exploited a probabilistic property inherent to the SLAM problem: If the entire camera motion {si} is known then the estimates of the positions of different... |

92 | Simultaneous localization and mapping with unknown data association using FastSLAM, in
- Montemerlo, Thrun
- 2003
(Show Context)
Citation Context ...able level. This paper describes a SLAM system using a single camera as the only sensor, with the specific aim of frame-rate operation with many landmarks. Estimates are maintained in a FastSLAM-style=-=[14]-=- particle filter. To our knowledge, this is the first use of such an approach in a monocular SLAM setting, presenting significant challenges: The FastSLAM particle filter model must be reconciled with... |

82 | Fusing points and lines for high performance tracking
- Rosten, Drummond
- 2005
(Show Context)
Citation Context ... depth of new landmarks, we estimate the inverse depth in the frame of first observation. Consider a newly observed landmark, L∗, selected automatically from the image (we use the feature detector of =-=[18]-=-). In the camera frame from which it is first observed, let its three-dimensional location be given by x∗ = ( x y z ) T (11) Instead of maintaining an estimate of these coordinates, we estimate the ca... |

79 | Structure from motion causally integrated over time
- Chiuso, Favaro, et al.
(Show Context)
Citation Context ...y the EKF. In the SFM literature, there has been significant success using the EKF for causal estimation – estimation depending only on observations up to the current time – with recursive algorithms =-=[2, 4, 9]-=-. In contrast to SFM approaches that rely on global nonlinear optimization, recursive estimation methods permit online operation, which is highly desirable for a SLAM system. Davison shows the feasibi... |

51 | Decoupled stochastic mapping, in
- Leonard, Feder
- 2002
(Show Context)
Citation Context ...ally initialized landmarks to help constrain camera pose. 2. Background 2.1. Scalable SLAM Extensive work has been undertaken in the robotics community to address the complexity of large scale mapping=-=[3, 7, 13, 23]-=-. Several approaches explicitly model the weak covariance between geographically distant landmarks by fully decorrelating their estimates in submaps. In a submap of bounded complexity, computation and... |

45 | Vision-based SLAM using the Rao-Blackwellised Particle Filter
- Sim, Elinas, et al.
- 2005
(Show Context)
Citation Context ...r to FastSLAM 2.0 to maintain pose and landmark estimates in SLAM. the landmark positions L j i 2.2. Vision SLAM with Particle Filters FastSLAM has been previously applied to vision-based SLAM by Sim =-=[20]-=-. However, Sim’s system uses a bottomup approach to SLAM, building a large database of feature descriptors into which features from novel views are matched to localize the robot. This approach preclud... |

41 |
Towards Constant Time SLAM using Postponement
- Knight, Davison, et al.
- 2001
(Show Context)
Citation Context ...time operation is desired. This level of map complexity allows localization and sparse mapping in a single room, but is not suited to very large areas or densely populated maps. Aggregated EKF updates=-=[8, 10]-=- allow efficient operation while observing a working set of landmarks, but full O(N 2 ) updates are still required when changing the working set. Since the number of landmarks N grows with time, O(N 2... |

30 | A semi-direct approach to structure from motion
- Jin, Favaro, et al.
(Show Context)
Citation Context ...y the EKF. In the SFM literature, there has been significant success using the EKF for causal estimation – estimation depending only on observations up to the current time – with recursive algorithms =-=[2, 4, 9]-=-. In contrast to SFM approaches that rely on global nonlinear optimization, recursive estimation methods permit online operation, which is highly desirable for a SLAM system. Davison shows the feasibi... |

29 | Real-time camera tracking using a particle filter
- Pupilli, Calway
- 2005
(Show Context)
Citation Context ... independence of landmarks given camera pose. Results are shown for only 33 landmarks, using 10,000 particles. The running time and scaling complexity of the system is not reported. Pupilli and Calway=-=[17]-=- use a particle cloud to represent camera pose hypotheses, while landmarks are represented communally. The focus of the work is on robust camera localization, so results with many landmarks are not sh... |

22 |
A practical 3D bearing-only SLAM algorithm
- Lemaire, Lacroix, et al.
(Show Context)
Citation Context ...KF. Lemaire et al. use a similar approach, but distribute depth hypotheses uniformly in inverse depth along the ray, as this corresponds to constant density of hypotheses when projected into the image=-=[12]-=-. As new measurements are made, Lemaire et al. repeatedly prune unlikely hypotheses until only one remains. A new landmark is initialized using the survivor hypothesis as a starting point. 5.1. Determ... |

20 | Application of Lie algebras to visual servoing
- Drummond, Cipolla
(Show Context)
Citation Context ...ics in SE(3) with the exponential map exp. To move a pose C by velocity µ over time δ, the pose is multiplied by the exponential of µ: Ct+δ = exp (δµ) · Ct (3) For details of this representation, see =-=[6, 9]-=-. We assume a constant velocity motion model for each camera hypothesis, similar to that given in [5]. During each time step, a continuous random velocity walk with zeromean white-noise acceleration o... |

7 |
time simultaneous localisation and mapping with a single camera
- Davison, “Real
- 2003
(Show Context)
Citation Context ...obal nonlinear optimization, recursive estimation methods permit online operation, which is highly desirable for a SLAM system. Davison shows the feasibility of real-time SLAM with a single camera in =-=[5]-=-, using the well-established EKF estimation framework. His system takes a top-down Bayesian estimation approach, searching for landmarks in image regions constrained by estimate uncertainty instead of... |

6 |
Bearing-only slam in indoor environments using a modified particle filter
- Kwok, Dissanayake
(Show Context)
Citation Context ...me of 11.9s. Furthermore, Sim’s system uses a stereo camera rig, which simplifies the observation model but does not match the flexibility and low footprint of a monocular system. Kwok and Dissanayake=-=[11]-=- use a modified particle filter to perform SLAM in a planar world by observing vertical edges with a camera. The system uses particle clouds to describe the probability distributions of landmarks in t... |

4 |
Efficient Simultaneously Localisation and Mapping in Large Environment
- Guivant
- 2002
(Show Context)
Citation Context ...ally initialized landmarks to help constrain camera pose. 2. Background 2.1. Scalable SLAM Extensive work has been undertaken in the robotics community to address the complexity of large scale mapping=-=[3, 7, 13, 23]-=-. Several approaches explicitly model the weak covariance between geographically distant landmarks by fully decorrelating their estimates in submaps. In a submap of bounded complexity, computation and... |

3 |
Decoupling localization and mapping in slam using compact relative maps
- Wang, Huang, et al.
- 2005
(Show Context)
Citation Context |