Creating Metric-topological Maps for Large-scale Monocular SLAM

Eduardo Fernández-Moral, Javier González-Jiménez and Vicente Arévalo

Departamento de Ingeniería de Sistemas y Automática, University of Málaga, Bulevard Louis Pasteur 35, Málaga, Spain

Keywords: Monocular SLAM, Metric-topological Map, Map Partitioning.

Abstract: In the last very few years, monocular SLAM approaches based on bundle adjustment are achieving amazing

results in terms of accuracy, computational efficiency, and density of the map. When such solutions are

applied on large scenarios it is crucial for the system scalability to maintain a map representation that

permits efficient map optimization and augmentation. In order to cope with such large maps, we present an

on-the-fly partitioning technique which allows abstraction from the metric map to operate more efficiently.

The result is a metric-topological arrangement where the areas with highly-connected observations are

grouped in submaps weakly interconnected to each other. This is accomplished by progressively cutting a

graph representation of the map, where the nodes are keyframes and the arcs between them represent their

shared observations. The experimental results indicate that the proposed approach improves the efficiency

of monocular SLAM and provides a metric-topological world representation suitable for other robotic tasks.

1 INTRODUCTION

Monocular SLAM is an appealing way of solving

the localization and mapping problem in mobile

robotics because cameras are inexpensive, compact,

easy to calibrate and consume low power. During

the last years monocular SLAM has advanced

notably with the use of parallel processing and

efficient algorithms for data association and map

optimization. It has made possible that current state-

of-the-art approaches can operate accurately in some

large scale scenarios, facilitating its application in a

wide range of areas such as augmented reality, scene

reconstruction and, particularly, mobile robotics.

The increasingly larger maps that are now

possible with monocular SLAM are fundamental to

cope with a wider range of real autonomous robotics

applications. Such ability to operate in large scale

brings the need of appropriate strategies for

managing the map. Applying abstraction (as humans

do) is an effective way of dealing with the huge

amount of detail present in large metric maps. The

result of such abstraction process is the so-called

metric-topological map, consisting of a two-layer

representation, one containing pure geometrical

information and a second one containing higher

level symbolic information (Thrun, 1998).

The benefit of a metric-topological arrangement

is twofold: on the one hand, it offers a natural

integration with symbolic planning that permits a

robot to reason about the world and to execute high

level tasks (Galindo et al., 2005; Segvic et al., 2009).

On the other hand, the efficiency and scalability of

the SLAM process itself are improved by limiting

the scope of localization and mapping to the region

of the environment where the robot is operating.

Also, loop closure and relocalisation can be more

efficiently solved using topological information

(Savelli and Kuipers, 2004; Angeli et al., 2009;

Fernández-Moral et al., 2013).

In this work, we present an online submapping

technique which creates a topological representation

of the world from the metric map being built by a

monocular SLAM technique. The key idea of our

proposal is to cluster in the same submap those

keyframes with higher observation overlap. This

presents some important advantages over other

approaches (as it will be explained latter on). The

generated map consists of a topological structure

composed of nodes representing local metric maps

and arcs representing relative geometric

transformations among the so-called submaps. In

this paper, we will focus on the benefits of such a

hybrid map for improving the efficiency and

scalability of conventional (metric) monocular

SLAM, concretely PTAM (Klein and Murray,

2007).

Next, we discuss some relevant related work and

Fernández-Moral E., Gonzalez-Jimenez J. and Arévalo V..

Creating Metric-topological Maps for Large-scale Monocular SLAM.

DOI: 10.5220/0004438900390047

In Proceedings of the 10th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2013), pages 39-47

ISBN: 978-989-8565-71-6

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

explain in detail the advantages of our approach.

We then describe our partitioning procedure and

show how it is combined with the SLAM process

(PTAM). The experiments and its results are

presented next, and finally, we expose the

conclusions of our work.

2 RELATED WORK

2.1 Construction of the Metric Map

Many solutions have been presented to build metric

maps with monocular SLAM since Davison

(Davison, 2003) presented the first real-time solution

for the problem in 2003. Two main strategies have

been applied since then: Bayesian filtering

(following the work of Davison) and Bundle

Adjustment (BA) on keyframes, as introduced in

(Klein and Murray, 2007). The latter represents the

base for the current state of the art since it allows

handling denser maps and generally offers a better

ratio accuracy/cost (Strasdat et al., 2010).

BA, traditionally used as an offline method for

Structure from Motion (SfM), is now widely used in

visual SLAM thanks to the introduction of parallel

processing and efficient algorithms which exploit

the sparse structure of the problem. Its application to

visual SLAM was inspired by real time visual

odometry and tracking (Nistér et al., 2005), where

the most recent camera poses where optimized to

achieve accurate localization. In such line, PTAM

selects keyframes and applies BA in a fixed size

window, around the last keyframe incorporated, to

obtain good metric maps and accurate localization.

Then, once the local optimization is performed, a

low priority global BA is run to improve the map

consistency. This approach is extended in (Holmes

et al., 2009) by combining it with relative bundle

adjustment - RBA - (Sibley et al., 2009), allowing

fixed-time, consistent exploration. An improvement

of the latter to exploit the problem' sparse structure

was recently presented by (Blanco et al., 2013).

The work of (Strasdat et al., 2011) is also related

to RBA, they propose a double window

optimization: a first window as in PTAM and a

second one including the periphery of the first to

improve consistency by optimizing a pose-graph.

Despite the impressive results obtained, such unique

map solution has intrinsic limitations for managing

maps of real large environments. To avoid such a

limitation, we propose a topological arrangement in

local metric maps.

2.2 Dividing the Map

Map division has been addressed in a number of

works. Some relevant examples are: the Atlas

framework (Newman et al., 2003), where a new

local map is started whenever localization performs

poorly in the current local map, or the hierarchical

SLAM presented in (Estrada et al., 2005), where

sensed features are integrated into the current local

map until a given number of them is reached.

However, none of these provides a mathematically

grounded solution based on the particular perception

of the scene.

In (Eade and Drummond, 2007), the map is

divided in nodes where the landmarks are

represented in a local coordinate frame and, these

landmarks are updated using an information filter.

This method uses the common features between

adjacent nodes to calculate their relative pose. A

different approach called Tectonic-SAM (Ni and

Dellaert, 2007) uses a “divide and conquer”

approach with locally optimized submaps in a

Smoothing and Mapping framework (SAM). This

approach is improved in (Ni and Dellaert, 2010) to

build a hierarchy of multiple-level submaps using

nested dissection.

Other works employ “graph cut” to divide the

map according to a measurable property of the map

observations. On that mathematical sound basis,

(Zivkovic et al., 2005) addresses the problem of

automatic construction of a hierarchical map from

images; (Blanco et al., 2008) generates metric-

topological maps using a range scanner, and

generalizes the approach for other sensors; and

(Rogers and Christensen, 2009) splits the map

within a Bayesian monocular SLAM framework to

reduce the problem complexity.

Our method, which also relies on graph cut,

differs from the above works in the way the graph is

constructed, which is specifically tailored for BA-

based monocular SLAM. Our approach resembles

also the stereo-SLAM framework of (Lim et al.,

2011) who divide the map keyframes into groups

(called segments) according to their geodesic

distances in the graph. On the contrary, our map

partitioning is independent of the keyframe

positions, and is only based on observations

acquired from the scene. Concretely, the map is split

where there are less shared observations, minimizing

the loss of information and therefore, enforcing the

coherency and consistency of the submaps.

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

3 MAP PARTITIONING

Splitting a map into locally metric consistent and

globally coherent regions provides some relevant

advantages for SLAM. Next, we explain the benefits

of such map structure (subsection 3.1), and describe

our proposal to obtain this metric-topological

arrangement of the map (subsection 3.2).

3.1 SLAM Improvements through

Hybrid Mapping

The advantages of applying a coherent map partition

in monocular SLAM are diverse: a) all the metric

data in each submap can be referred to a local

coordinate system, what reduces error accumulation

and numerical instability; b) localization can be

achieved more efficiently since only those map

points in the nearer regions are reprojected to

estimate the camera position; c) this map structure

permits to approximate the global BA by the

individual optimization of the different submaps,

thus reducing the computational cost of the

optimization process. This last advantage is of

special relevance due to the demanding nature of

BA, whose complexity ranges from linear to cubic in

the number of keyframes depending on the particular

point-keyframe structure (Konolige, 2010). Next, we

explain the details of this approximation for the

global optimization.

Having a map of n landmarks obtained from

observations at m keyframes, bundle adjustment can

be expressed as





ijijij

)),,((min xbaQ

(1)

where

 d(x,x’) denotes the Euclidean distance between

the image points represented by vectors x and x’,

 a

is the pose of camera at keyframe j and b

the

position of landmark i,

 Q(a

) is the predicted projection of landmark i

on the image associated to keyframe j,

 x

represents the observation of the i-th 3D

landmark on the image of keyframe j and,

 v

stands for a binary variable that equals 1 if

landmark i is visible in keyframe j and 0

otherwise.

Let’s now consider that the map is divided into N

submaps, each submap, say k, containing m

keyframes and n

landmarks, with k = {1,…N}.

Then, (1) can be rewritten as

 

  













11 1 1

)),,((min xbaQ

(2)

where the combination of subscript i and superscript

k refers to the i-th landmark of the k-th submap (e.g.,

), and similarly l over j refers to the j-th keyframe

of the l-th submap (e.g., a

). Taking into account the

observations shared between submaps, this

expression can be written as































111

)),,((

)),,((min

xbaQ

(3)

where the term A stands for the reprojection error of

those landmarks observed from keyframes of

different submaps and the term B corresponds to the

reprojection error of those landmarks observed form

keyframes within the same submap. Both concepts

are illustrated in figure 1.b. The first establishes the

inter-connection between submaps which is

represented by arcs connecting keyframes of

different submaps (e.g. arc linking KF-2 and KF-11)

and the second sets the intra-connection of the

submap which includes the submaps’ inner arcs (e.g.

arc linking KF-1 and KF-2).

If we are able to divide the map in such a way

that the different submaps have few common

observations, and assuming that the reprojection

errors are independent of the map division, then A

becomes negligible with respect to B. Thus, the

global optimization can be approximated by

















ijijij

111

)),,((min xbaQ

(4)

This approximation is equivalent to optimize

each submap independently, which leads to a

significant reduction of computational burden. In

fact, this approximation is equivalent to the original

expression (1) when there are no connections

between submaps.

CreatingMetric-topologicalMapsforLarge-scaleMonocularSLAM

Figure 1: a) Common observations between two keyframes. This is used to calculate the Sensed Space Overlap (SSO) (see

equation 5). b) Graph-representation of the map where each node represents a keyframe and the arcs are weighed with the

SSO calculated between keyframes (thicker arcs mean higher SSO). c) Example of SSO matrix, in which the brightness of

the element ij represents the SSO between the keyframes i and j.

3.2 Map Partitioning Method

The approach proposed here to divide the map into

coherent regions consists in grouping together those

keyframes that observe the same features from the

environment. For that, we consider the map as a

graph whose nodes represent keyframes and the

weight of the arcs are a measure of the common

observations between them. There are two critical

issues in this partitioning approach: first, the

computation of the arc weights; and second, the

criterion adopted to perform the partition itself.

As for the first, the arc weights are assigned

according to the Sensed-Space-Overlap (SSO),

following our previous work (Blanco et al., 2006),

particularized for landmark observations. This

simple but effective measure represents the

information shared by two keyframes. It is

calculated with the relation between the number of

common landmark observations and the total

number of landmarks observed in both keyframes

(see figure 1.a). This is expressed as









vvvv

kfkf ),SSO(

(5)

where v

and v

, similarly to the definitions of the

previous section, are binary variables that equal 1 if

landmark i is observed in the keyframes kf

and kf

respectively.

Regarding the criterion for partitioning the

graph, we follow previous works (Zivkovic et al.,

2005; Blanco et al., 2008; Rogers and Christensen,

2009) that apply the minimum normalized-cut (min-

Ncut), originally introduced in (Shi and Malik,

2000). The min-Ncut has the desirable property of

generating balanced clusters of highly

interconnected nodes, in our case clusters of

keyframes that cover the same part of the

environment. Figure 1 illustrates this concept: figure

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

1.a shows the common observations in a pair of

keyframes whose arc weight is calculated with the

SSO (see Eq. 5), and figure 1.b shows a map

division into three submaps as produced by the min-

Ncut procedure. Notice that the pairs of keyframes

with higher SSO (thicker arcs) are grouped together.

Figure 1.c shows the symmetrical SSO matrix

corresponding to a different, larger map, where the

keyframes are arranged according the min-Ncut to

give rise to three groups of keyframes or submaps

(matrix blocks).

It is important to notice that, in order to

guarantee a scalable system when applying map

partitioning to visual SLAM, the size of the submaps

(i.e. number of keyframes) must be kept bounded.

This requirement is not demonstrated

mathematically here but it is intuitive to see that as

the camera explores new parts of the scene, the new

keyframes will have low SSO values (if any) with

distant ones in the map. Therefore, the min-NCut

will produce new partitions when the system

explores unobserved regions of the environment.

This can be more clearly understood with the

following example: let’s consider the case where

there are features that are always observed (e.g. the

horizon when travelling by train, or when zooming

in the scene, or traversing a corridor with the camera

pointing in the movement direction) as the new

keyframes are selected, they will introduce new

features and therefore will reduce the minimum

normalized-cut, resulting in the eventual partition of

the map. The last two examples represent another

advantage of our partition method, which produces

natural multi-scale maps when the camera zooms.

This insight is supported by all the experiments we

have carried out during this work.

4 COMBINATION OF MAP

PARTITIONING AND PTAM

This section outlines the combination of our

partition procedure and Parallel Tracking and

Mapping (PTAM) (Klein and Murray, 2007). PTAM

is a monocular SLAM algorithm which performs

online BA on keyframes, separating the tracking and

mapping stages in two different threads to permit

efficient real-time execution. This technique requires

an initial map before it starts working autonomously.

Such initial map is acquired with a Structure from

Motion procedure that involves user intervention to

select two views with sufficient parallax. Once the

Figure 2: Tracking and mapping threads of PTAM. Blue

boxes correspond to the embedded stages to perform the

map partitioning.

initial map has been created, the system analyses the

images retrieved by the camera to self-localize in the

map, while the map is continuously optimized and

augmented with new keyframes and landmarks.

Such keyframes are selected according to some

simple heuristics (see (Klein and Murray, 2007) for

more details), and new landmarks are extracted

through epipolar search between each new keyframe

and its nearest keyframe in the map.

Figure 2 shows a scheme of the proposed

partitioning method interacting with PTAM. Our

submapping procedure takes action in both of PTAM

threads. In the tracking thread, it selects the current

submap and the nearest keyframe to the estimated

pose after a new image is analyzed. In the mapping

thread, after a new keyframe is selected and new

landmarks are detected in it, the SSO is evaluated

with respect to all the keyframes of the vicinity,

which includes all the submaps directly connected to

the current submap (see figure 3). The partitioning

procedure comes into play after the SSO has been

updated, then, the min-Ncut is evaluated, and if it

results in a different partition, the map is rearranged.

This partitioning method is applied dynamically as

the map enlarges and may create new submaps as

well as merge existing submaps to maintain

coherency by grouping keyframes with high overlap.

The partitioning process, including SSO

computation, min-NCut evaluation and map

rearrangement depends on the number of keyframes

and landmarks in the vicinity, taking up to 100 ms.

in our experiments, which supposes a short time in

comparison with the map optimization time.

CreatingMetric-topologicalMapsforLarge-scaleMonocularSLAM

Figure 3: Topological representation of the concept of

submap vicinity.

5 EXPERIMENTAL RESULTS

In this section we present some experiments which

show the advantages, in terms of efficiency and

scalability, of using the proposed metric-topological

arrangement of the map instead of a single metric

map. The experiments have been carried out using a

Philips SPC640NC webcam, connected by USB to a

linux-based laptop with an Intel Core2 Duo 2.4 GHz

processor, 2Gb of memory and a nVidia GeForce-

9400 graphics card.

A first experiment is aimed to illustrate the

increase of efficiency in localization at frame rate.

For that, we compare the time needed to project map

points into the current frame with and without

partitioning as the map grows. Both tests have been

performed in the same environment, building maps

composed of about 45000 points and 1000

keyframes, distributed in 52 submaps for the

partitioning case. Figure 4 shows that the time with a

unique map grows linearly with the number of map

points, whereas with submapping, this time is

bounded since only those points in submaps close to

the camera are evaluated. This improvement in

efficiency becomes more relevant when the map

grows nonstop (note that this process is performed

with each new frame captured by the camera).

The goal of a second experiment is to quantify

the efficiency in the global optimization of the map

with our submapping approximation. For that, we

have run BA offline after every new keyframe is

selected from a recorded video (that is, sequential

SfM), measuring the times of each BA completion

with and without partitioning. At the end of these

tests, the maps created were composed of about

22000 points and 400 keyframes, distributed in 9

submaps for the partitioning case. In order to

compare both alternatives in the same conditions, we

have included the time of partition management in

the BA time for the partitioning test. Figure 5 shows

the optimization times vs. the number of keyframes

of the whole map for both cases. As expected, for

the case without partitioning, the computational cost

follows an increasing polynomial trend with the

number of keyframes. Conversely, when applying

map partitioning, the computational burden is

bounded since the BA is applied only on the current

submap. For this case, we can observe some abrupt

changes in the cost which are produced when the

reference submap (the one where the system is

localized) switches to a neighbor of different size.

Figures 6.a and 6.b show the maps built with both

alternatives (different colors represent different

submaps in 6.b). We can verify visually their high

similarity, and their good alignment, as a result of

the continuous optimization previous to the map

partition.

Additionally, we are interested in comparing the

accuracy of the generated metric map. Due to the

lack of a reliable metric to evaluate the map’s

quality, we have compared visually the different

maps considering as ground truth the map obtained

offline in the previous experiment (figure 6.a),

which is the most accurate we can get. In the map

obtained with PTAM (figure 6.c), we can appreciate

some regions with depth errors and many outliers

(e.g. landmarks detected behind physical walls).

These inconsistencies are consequence of the

premature interruption of global BA that happens

when a new keyframe is selected, what leads to data

association errors and the subsequent accuracy

decrease with the map size. On the contrary, the map

obtained with our approach (figure 6.d) presents no

inconsistencies and considerably less outliers than

the unique map solution (figure 6.c). This results

from the higher efficiency of the submap local

optimization, which optimizes regions with highly

correlated observations to produce locally accurate

submaps.

The results shown in this section have been

supported in several tests performed under different

conditions: exploring different rooms, re-visiting

previous maps, traversing a corridor, zooming to get

more detail of the scene, etc. The reader may refer to

http://mapir.isa.uma.es/eFernández/partition for a

video that illustrates the operation of our

submapping approach with PTAM in different

environments.

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

Figure 4: Map projection time for localization with and

without map partitioning.

Figure 5: Bundle adjustment computation time (offline)

with and without partitioning.

Figure 6: Top view of maps generated in our experiments. All the maps are composed of more than 400 keyframes and

22.000 landmarks. The different colors in b) and d) represent different sub-maps.

6 CONCLUSIONS

This article presents an online submapping method

which transforms a metric map into a metric-

topological arrangement of it. This hybrid metric-

topological structure improves the scalability of

monocular SLAM in two aspects: first, the system

rules out unnecessary metric information to perform

more efficiently; second, it permits to use an

CreatingMetric-topologicalMapsforLarge-scaleMonocularSLAM

approximation of BA to reduce computational cost

while maintaining map consistency. Besides, the

topological arrangement of the map is useful for

other tasks, as loop closure, global localization or

navigation. Experiments have demonstrated the

potential of our approach to obtain efficient map

representation in large environments. Future work

will focus on exploiting the topological structure of

the map for tasks as loop closure and relocalisation.

ACKNOWLEDGEMENTS

This work has been supported by the project

“TAROTH: New developments toward a robot at

home”, funded by the Spanish Government and the

“European Regional Development Fund ERDF”

under contract DPI2011-25483.

REFERENCES

Angeli, A., Doncieux, S., Meyer, J.-A., Filliat, D., 2009.

"Visual Topological SLAM and Global Localization",

In IEEE International Conference on Robotics and

Automation.

Blanco, J. L., Fernández-Madrigal, J. A., González, J.,

2008. Toward a unified bayesian approach to hybrid

metric-topological SLAM. IEEE Transactions on

Robotics and Automation, 24(2):259-270.

Blanco, J. L., González, J., Fernández-Madrigal, J. A.,

2006. “Consistent observation grouping for generating

metric-topological maps that improves robot

localization”. In IEEE International Conference on

Robotics and Automation, pp. 818–823.

Blanco, J. L., González-Jiménez, J., Fernández-Madrigal,

J. A., 2013. "Sparser Relative Bundle Adjustment

(SRBA): constant-time maintenance and local

optimization of arbitrarily large maps", In IEEE

International Conference on Robotics and

Automation.

Davison, A. J, 2003. “Real-time simultaneous localisation

and mapping with a single camera,” In Proceedings of

the International Conference on Computer Vision.

Eade, E., Drummond, T., 2007. “Monocular slam as a

graph of coalesced observations”. In Proceedings of

the International Conference on Computer Vision.

Estrada, C., Neira, J., Tardos, J., 2005. “Hierarchical slam:

Real-time accurate mapping of large environments”.

IEEE Transactions on Robotics, vol. 21, no. 4, p. 588-

596.

Fernández-Moral, E., Mayol-Cuevas, W., Arévalo, V.,

González-Jiménez, J., 2013. "Fast place recognition

with plane-based maps", In Proceedings of the IEEE

International Conference on Robotics and

Automation.

Galindo, C., Saffiotti, A., Coradeschi, S., Buschka, P.,

Fernández-Madrigal, J. A., J. González, 2005. “Multi-

hierarchical semantic maps for mobile robotics,” In

Proceedings of the IEEE/RSJ International

Conference on Intelligent Robots and Systems, pp.

2278–2283.

Holmes, S. A., Sibley, G., Klein, G., Murray, D. W., 2009.

”A relative frame representation for fixed-time bundle

adjustment in monocular SFM”. In Proceedings IEEE

International Conference on Robotics and

Automation.

Klein, G., Murray, D. W, 2007. “Parallel tracking and

mapping for small AR workspaces”. In Proceedings of

the International Symposium on Mixed and

Augmented Reality.

Konolige, K., 2010. “Sparse sparse bundle adjustment”. In

Proceedings of the British Machine Vision

Conference.

Lim, J., Frahm, J. M., Pollefeys, M., 2011. Online

environment mapping. In Proceedings of the IEEE

Conference on Computer Vision and Pattern

Recognition.

Newman, P., Leonard, J., Soika, M., Feiten, W., Teller, S.,

2003. “An atlas framework for scalable mapping”. In

Proceedings of the IEEE International Conference on

Robotics and Automation, vol. 2, pp. 1899–1906.

Ni, K., Dellaert, F., 2010. “Multi-level submap based slam

using nested dissection”. In

IEEE/RSJ International

Conference on Intelligent Robots and Systems.

Ni, K., Steedly, D., Dellaert, F., 2007. “Tectonic SAM:

Exact, out-of-core, submap-based SLAM,” in IEEE

International Conference on Robotics and

Automation.

Nistér, D., Naroditsky, O., Bergen, J.R, 2005. “Visual

odometry”. In Proc. IEEE International Conference on

Computer Vision and Pattern Recognition, pages 652–

659.

Rogers, J. G., Christensen, H. I., 2009. “Normalized graph

cuts for visual slam”. In IEEE/RSJ International

Conference on Intelligent Robots and Systems.

Savelli, F., Kuipers, B., 2004. “Loop-Closing and

Planarity in Topological Map-Building,” In

Proceedings of the IEEE/RSJ International

Conference on Intelligent Robots and Systems, vol. 2,

pp. 1511–1517.

Segvic, S., Remazeilles, A., Diosi, A. and Chaumette, F.,

2009. "A mapping and localization framework for

scalable appearance-based navigation". Computer

Vision and Image Understanding 113(2): 172-187.

Shi, J., Malik, J., 2000. “Normalized cuts and image

segmentation”. IEEE Transactions on Pattern

Analysis and Machine Intelligence, vol. 22, no. 8, p.

888-905.

Sibley, G., Mei, C., Reid, I., Newman, P., 2009. “Adaptive

relative bundle adjustment”. In Robotics Science and

Systems.

Strasdat, H., Davison, A. J., Montiel, J. M. M., Konolige,

K., 2011. “Double Window Optimisation for Constant

Time Visual SLAM”. In IEEE International

Conference on Computer Vision.

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

Strasdat, H., Montiel, J. M. M., Davison, A. J., 2010.

“Real-time monocular slam: Why filter?”. In IEEE

International Conference on Robotics and

Automation.

Thrun, S., 1998. “Learning Metric-Topological Maps for

Indoor Mobile Robot Navigation,” Artificial

Intelligence, v.99, no.1, pp. 21-71.

Zivkovic, Z., Bakker, B., Krose, B., 2005. “Hierarchical

map building using visual landmarks and geometric

constraints”. In Proceedings of the IEEE/RSJ

International Conference on Intelligent Robots and

Systems, p. 2480-2485.

CreatingMetric-topologicalMapsforLarge-scaleMonocularSLAM