Solving the Indoor SLAM Problem for a Low-Cost Robot using

Sensor Data Fusion and Autonomous Feature-based Exploration

Luciano Buonocore

, Cairo Lúcio Nascimento Júnior

and Areolino de Almeida Neto

Universidade Federal do Maranhão, Av. dos Portugueses S/N, São Luis, Maranhão, Brazil

Instituto Tecnológico de Aeronáutica, São José dos Campos, Brazil

Keywords: Sensor Data Fusion, Feature-based Autonomous Exploration, Visual Sensor, Low-Cost Robot, SLAM.

Abstract: This article is concerned with the solution of the SLAM (Simultaneous Localization And Mapping) problem

in an indoor environment using a low-cost mobile robot that autonomously explores the environment. The

robot was constructed with a distance measurement subsystem composed of three types of sensors: a

wireless webcam with a laser pointer (a visual sensor), two infrared sensors and an ultrasonic TOF (time-of-

flight) sensor. Firstly, an algorithm that requires a small computational load is used to fuse the noisy raw

data acquired by these sensors and generate the environment features. These features are then used by a

particle filter to solve the SLAM problem. An autonomous feature-based exploration algorithm was

implemented and is also presented. The results obtained in the experiments carried out in two small indoor

environments show that the estimated environment map generated when the robot uses the autonomous

exploration algorithm is very similar to the one generated when the robot poses were manually chosen.

1 INTRODUCTION

Nowadays it is highly desirable that a mobile robot

is able to navigate in an unknown environment

autonomously, which means that the robot needs to

construct a map of the environment and at the same

time locate itself within this map, a problem

commonly referred as SLAM (Simultaneous

Localization And Mapping) in the robotics literature.

Three software modules, which run in a

cooperative form, were used (Figure 1): a) a sensor

data fusion algorithm; b) an autonomous feature-

based exploration algorithm; and c) a version of

particle filter (Rao-Blackwellized). The two first

modules were developed by the authors and are the

main contributions of this article.

The robot, which was designed and built by the

authors of this article, named SLAMVITA (final

cost around US $ 2,100.00) employs three types of

sensors: a wireless webcam with a laser pointer (a

visual sensor), two infrared sensors and an ultrasonic

TOF (time-of-flight) sensor or sonar. An algorithm

was developed to transform the noisy data acquired

by these sensors into features that geometrically

represent the environment (walls).

A particle filter algorithm, known as Rao-

Figure 1: Algorithm modules to solve the indoor SLAM

problem.

Blackwellized or FastSLAM (Huge and Bailey,

2006), was implemented based on a proposal by

(Thrun et al., 2005) which was modified to use the

odometry motion model, such that the outputs of the

encoders were computed as control information.

This algorithm simultaneously estimate the robot

poses and the environment map in solution of

SLAM problem.

The autonomous exploration algorithm is

structured as rules to perform two basic tasks: a) to

decide the best target to explore in the environment

in an incremental way; and b) to decide when to

finish the exploration task.

This article is organized as follows. Section 2

introduces the SLAM problem and the approaches

used to solve it. In Section 3 the proposed sensor

data fusion algorithm is presented and shows the

407

Buonocore L., Lúcio Nascimento Júnior C. and de Almeida Neto A..

Solving the Indoor SLAM Problem for a Low-Cost Robot using Sensor Data Fusion and Autonomous Feature-based Exploration.

DOI: 10.5220/0004000504070414

In Proceedings of the 9th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2012), pages 407-414

ISBN: 978-989-8565-22-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

results of a simple experiment which was designed

to measure its accuracy for environment mapping.

Section 4 explains the autonomous exploration

algorithm. Section 5 presents the experiment results

obtained by the solution for the SLAM problem with

and without using the proposed autonomous

exploration algorithm in a small indoor environment.

The same Section shows that the proposed solution

fails if the raw data of any two SLAMVITA robot

sensors are used by the fusion algorithm.

2 SLAM PROBLEM OVERVIEW

Some approaches employed to solve the SLAM

problem were obtained using different techniques,

such as fuzzy logic (Huge, 2001; Aguirre and

Ganzález, 2002), artificial neural networks (Thrun,

1993), and the Dempster-Shafer theory

(Milisavljević, Bloch and Acheroy, 2008). However,

most of the literature employs the so-called

probabilistic SLAM solutions (Thrun, Burgard and

Fox, 2005), using different combinations of sensors,

such as laser scanner and CCD camera (Castellanos,

et al., 1998) or two sonars and six infrared sensors

(Vazquez and Malcolm, 2005). Many other sensor

combinations may be found in the literature.

Figure 2 shows a graphic representation of the

two main forms of probabilistic SLAM problems

(Thrun et al., 2005): online SLAM and full SLAM.

In the solution of the SLAM problem the robot poses

and environment map are unknown estimated

variables (X and m, respectively) that must be

computed at every time step t, and the accuracy of

one variable is important to best estimate the other.

This fact is referred in the literature as the “chicken

and egg” problem.

This article uses a version of a particle filter

known as Rao-Blackwellized or FastSLAM, where

samples (particles) are employed to hold hypothesis

of the robot poses along the EKF (Extended Kalman

Filter) to estimate each feature extracted by the

proposed sensor data fusion algorithm. Moreover,

the choice for FastSLAM version 1.0, instead of

version 2.0, is due to the high quality odometry

Figure 2: Forms of SLAM problems.

(Thrun et al., 2005) shown by the SLAMVITA robot

in the experiments.

3 PROPOSED SENSOR DATA

FUSION ALGORITHM

A single sensor hardly ever can provide enough

information about the environment because of its

limited field-of-view. The solution to overcome such

shortcomings requires the use of two or more

sensors whose information must be fused. Three-

level architecture for robot sensor data fusion is

presented by (Visser, 1999) in which three methods

of data fusion may be applied: a) cooperative; b)

competitive; and c) complementary.

The proposed sensor data fusion algorithm only

uses competitive and complementary methods. In a

competitive fusion each sensor produces its own

estimated parameters, which are combined to

generate a single set of parameters (in this case, the

distance and the angle of the measured object). An

example of competitive fusion occurs when the

visual and infrared sensors measure the same object

in the environment. Complementary fusion is

applied when each sensor has only partial

information about the environment. It aims to

overcome the incompleteness of sensors being a

classical example the array of several sonars

attached around a mobile robot.

Figure 3: Measurement parameters (d, Ɵ) provided by the

SLAMVITA sensors: (a) visual, (b) sonar and (c) infrared.

ICINCO 2012 - 9th International Conference on Informatics in Control, Automation and Robotics

408

The SLAMVITA sensors are assembled on a

panning turret located at top of the robot (Figure 3).

The turret and the robot vertical axis are aligned and

the turret can be turned around its axis to execute an

180° scan procedure.

The parameters supplied by the sensors are the

distance (d) and angle (Ɵ) of the object measured

(details in Figure 3). These sensors were chosen in

function of the complementary information provided

by them. Due to their operating principles, the visual

and infrared sensors provide reliable directional

information (angle of the object), while the sonar is

mainly a distance sensor, because of its accuracy in

distance measurement (Ivanjko et al., 2009; Vazques

and Malcolm, 2005).

The visual sensor employed in SLAMVITA

robot uses active triangulation with calibration

targets proposed by (Nguyen and Blackburn, 1995)

and has never been used in the robotic literature for

mapping purpose. The operating principle (laser

center on image calculated in subpixel resolutions)

and experimental results performed in distance

measurements by the visual sensor are presented in

(Buonocore et al., 2010).

The proposed sensor data fusion algorithm is

informally described next to facilitate its

comprehension, which flowchart is shown in Figure

4. It is based on the following considerations:

 It is assumed that the robot environment is

composed of walls (they are always orthogonal

Figure 4: Flowchart of the proposed sensor data fusion

algorithm.

to the floor but not necessarily to each other),

and these walls can have openings (such as

doors). Most structured indoor environments are

composed by walls, doors, etc. that may be used

as line segments into a compact map

representation (Yap and Shelton, 2009).

 The occupancy grid (OG) and feature-based

mapping techniques are employed to represent

the environment in two dimensions.

 The RANSAC algorithm (Zuliani, 2009) is used

to extract straight line segments from the

directional (visual and IR) maps.

 The sonar measurements available in OG map

(OGs) and in its vector raw data are used to

eliminate noises in the directional OG maps and

to adjust the feature distances in the directional

fused feature-base map, respectively.

 The distance parameters in the visual and IR

sensors are considered to have the same weights

to compute the straight line segments to be fused.

The input data (ID) shown in Figure 4 are the

distance/angle contained in each sensor raw data

acquired in an 180

scan, at each robot pose, with

1.8

angular resolution. The algorithm generates the

features extracted (output data, OD) as parameter-

zed lines. The proposed fusion algorithm is com-

posed by 4 specific algorithms and 3 fusion rules.

These rules (R1, R2 and R3) implement the data

fusion method while the specific algorithms (A1 to

A4) prepare the data to feed the rules. The pseudo-

codes for these procedures are available at ftp:

//labattmot.ele.ita.br/ele/luciano/My_Publications.

The basic concept in R1 rule is as follows: no

object detected by the sonar cone must be perceived

by the other two directional sensors. The application

of R1 rule justifies the use of the OG representation,

where the inverse sonar model is employed (Thrun

et al., 2005). Due to the surface reflection that is not

normal to the sonar acoustic wave, the sonar raw

data must be pre-processed before being employed

to remove these wrong data, keeping only those ones

that mainly reveal the true RCD (Region of Constant

Depth), reporting for walls in the environment

(Pandey et al., 2007).

To extract line segments from directional maps,

the proposed algorithm converts the OGv and OGir

internal representation (after applying R1 rule) to x-y

coordinates, which are held in separated vectors.

Then, the RANSAC method is used in A4 algorithm

to extract straight line segments (features) from

those vectors, one map at a time. For each

directional map, the A4 algorithm performs an

interpretation on the data vector to break it in

Solving the Indoor SLAM Problem for a Low-Cost Robot using Sensor Data Fusion and Autonomous Feature-based

Exploration

409

Figure 5: Comparison between OG maps computed from

each sensor raw data and one generated by the proposed

fusion algorithm.

segments that have significant variance in its data

that point to different straight line slope or also gaps

that may appear between line segments

(environment openings). All vector segments are

submitted to RANSAC trying to extract straight

lines that represent the section features of the

processing directional map.

The line segments extracted for each map are

combined in R2 rule, applying competitive or

complementary method. The competitive fusion is

selected if both sensors have data at specific angular

area. Otherwise, the complementary fusion takes

place, such as the case in which the visual sensor is

unable to define the laser spot at the camera image.

The R3 rule adjusts, if possible, the feature

distances of the fused directional map (after

applying R2 rule), searching for RCD in the angular

area that contains the normal straight line of each

feature in the map. Whenever the RCD is found, the

feature distance is adjusted using the sonar

measurement (RCD distance).

Figure 5 shows the geometrical representation of

the actual environment (photo), which in comparison

with each OG sensor map is more consistent and

noise free. The error obtained in this mapping

experiment was less than 1.5%. It is important to

notice that the map constructed in this experiment is

represented as full OG maps, instead of feature-base

ones. However, in SLAM applications, the features

for each robot pose must be processed and the

environment representation outputted by the

algorithm changed to feature-based map because of

its computational cost in the particle filter.

4 AUTONOMOUS

EXPLORATION ALGORITHM

The robot ability to explore an environment without

human intervention ensures its real autonomy. The

proposed sensor data fusion algorithm is nicely

complemented by an autonomous exploration

algorithm that uses the same environment map

representation. The autonomous exploration

algorithm developed in this article, from now on

denoted by AEA, is based on concepts presented by

(Newman, Bosse and Leonard, 2003) where

hallways were mapped using two basic action

criteria: a) goal generation and b) goal selection.

Other concepts, such as visibility evaluation, were

regarded in the AEA implementation.

It is important to note if no feature is extracted

for a specific robot pose, the uncertainty in the

estimated robot pose by the particle filter increases.

The AEA deals with this case by using incremental

exploration, avoiding the space-free analysis

approach (e.g., Voronoi diagrams).

Figure 6 shows the flowchart of the proposed

AEA. In the proposal presented in this article, only

local context is relevant for the goal chosen, while in

(Newman, Bosse and Leonard, 2003) local and

global context are considered to goal selection. In

the proposed AEA switching from local to global

context is necessary only to define the goal that is

located in the environment opening, which allows

the robot to reach the selected global goal. This

modification is necessary because the SLAMVITA

robot must acquire the sensor measurements when it

is stopped, after executing the movements between

consecutive poses. Local to global context switches,

such as presented in (Newman, Bosse and Leonard,

2003) would cause a large increase in the time

necessary to finish the exploration task. In other

words, the goals reached by the SLAMVITA robot

are localized either locally (local context) or in an

environment opening (global context evaluated).

Besides the definition of the best goal to explore

in the environment, another important resource that

the AEA performs is the decision of finishing the

mapping task.

ICINCO 2012 - 9th International Conference on Informatics in Control, Automation and Robotics

410

5 EXPERIMENTS AND RESULTS

Several experiments in distinct small environments

were carried out, with or without AEA. In this

Section, two of the experiments that use the same

environment are presented, differing in the

exploration forms (manual or autonomous). Figure 7

shows the environment (8.8 m x 2.80 m or 24.64m

where the two loops show the trajectory traveled by

the robot when the exploration was manually

planned. In this environment and with the robot

starting at same pose, the second experiment was

performed using AEA.

Figure 6: Flowchart of the proposed featured-based

Autonomous Exploration Algorithm (AEA).

Figure 7: Indoor environment for the SLAM experiments.

Figure 8(a) shows the estimated map with the

robot poses manually planned, while Figure 8(b)

shows the estimated map using AEA. Table I

informs the quantitative absolute difference between

the estimated and real robot poses relative to both

experiments. Some important considerations that can

be mentioned in the experiment results, shown in

Figure 8(a, b) and Table I are:

Figure 8: SLAM experiments with the robot using: (a)

manually planned movements; and (b) autonomous

exploration.

a) The absolute differences between the estimated

and real robot poses (ΔX, ΔY and ΔƟ) in both

experiments are kept small, less than 2%.

However, in some robot poses, mainly in the

experiment with AEA, the difference increased

up to 3.5% (e.g., poses 32 and 33). The reason

for this is the accuracy of the features acquired

in some estimated poses where the robot made

its measurements, that is, they were not

associated with the proposed AEA.

b) The number of poses where the measures were

taken is more than 50%, when using AEA (36

poses). Six robot poses (from 26 to 31 in Figure

8(b)) can be avoided, just evaluating the

opening that has potential exploration to justify

the movement of the robot to them. It is the

case of two goals selected at environment

opening in the experiment with AEA that led

the robot to move in their directions.

c) The path followed by the robot in the

experiment with AEA is not based in loop

closing. The criterion in AEA is based just in

goal scoring (locally and environment opening,

the last using a global context).

d) The estimated maps for both experiments are

consistent with the real environment contour for

navigation purposes. Some segments are out of

Solving the Indoor SLAM Problem for a Low-Cost Robot using Sensor Data Fusion and Autonomous Feature-based

Exploration

411

Table I: Absolute differences between estimated and real

robot poses (ΔX and ΔY in cm, ΔƟ in degree).

Without AEA

With AEA

POS

ΔX

ΔY

ΔƟ

ΔX

ΔY

ΔƟ

0.00

-1.00

-0.32

-2.00

0.51

0.15

1.85

-2.48

4.52

0.00

0.84

-2.72

1.47

-1.01

1.36

0.00

0.84

-2.72

1.42

-1.65

0.44

0.00

-3.29

-5.23

1.00

-3.33

0.81

-2.00

-0.21

0.85

2.98

1.85

-1.51

-3.04

-0.21

0.85

3.58

6.69

-1.63

0.00

-0.51

3.83

-1.85

2.86

-2.26

0.00

1.94

1.97

0.24

2.53

-0.20

-2.00

0.94

2.97

-0.24

2.94

1.92

3.00

0.28

2.96

-2.15

0.00

0.01

0.00

0.64

2.47

0.00

-0.67

1.99

0.00

0.64

2.47

0.00

-0.19

-0.70

-2.00

0.56

2.57

0.71

0.97

-4.01

0.00

0.56

2.57

0.10

2.86

-3.73

0.00

0.92

0.17

0.00

4.72

-2.97

0.00

0.92

0.17

1.00

4.50

-2.43

-1.00

2.36

3.77

3.03

0.07

0.05

0.43

2.36

3.77

1.77

-1.09

-0.04

-2.00

-1.76

4.84

2.99

1.36

-3.61

-1.84

-1.76

4.84

2.93

-2.39

2.91

3.12

-2.39

2.91

0.27

5.02

-1.83

0.99

6.35

-1.50

-0.06

4.44

-0.79

0.80

3.51

-2.20

-0.80

4.07

-4.05

3.39

1.62

3.22

3.13

-0.58

1.86

4.97

2.19

10.89

2.34

2.19

10.89

0.91

0.94

9.74

0.00

-1.43

7.67

0.99

-1.43

7.67

0.72

the actual environment area and others can be

viewed more length than the respective walls

inside the environment (Figure 8(b)). Once

again, this is not associated to employing AEA

and a possible solution to remove this mistakes

is evaluating feature crossing (inside and on

borders of the environment) based on the

estimated robot poses. If the segment

augmented in the feature crossing cannot be

perceived in any estimated robot pose (border

situations) or even if it occludes the previous

ones that were viewed in other estimated robot

poses (feature inside the environment), the

feature excess may be removed accordingly.

Although the estimated map without AEA

(Figure 8(a)) presents to be closer to the actual

environment than the one with AEA (Figure8(b)),

the fundamental reason to use the SLAM solution

with AEA approach is the robot´s autonomy,

avoiding human interventions as much as possible.

With some data produced in the experiment

without AEA (robot commands and sensor raw data

acquired in 22 position), one of the three sensors

available in SLAMVITA was “withdrawn”, just not

involving its data in the fusion algorithm. So, the

data (commands and raw data of the two “remained”

sensors) are used to verify if it is possible to validate

the solution to SLAM problem.

The experiments conducted for all combinations

of two SLAMVITA sensors with the total features

extracted by the proposed fusion algorithm

considering the robot real poses are shown in Figure

9(a), while the estimated map outputted by the

particle filter after the first six poses are presented in

Figure 9: (a) The total feature acquired by two

SLAMVITA sensors for all real robot poses; and (b) the

estimated map by the filter until robot pose #6 (1-visual

and sonar; 2- visual and infrared; and 3-infrared and

sonar). (b) Just few poses were sufficient to show that the

filter diverge for all cases.

When using the infrared and visual sensors (case

2, and so, not applying R1 and R3 rules) the overall

features appears with some distortions of the real

environment (Figure 9(a)), but with less feature

mistakes related to the others sensors combinations.

However, in all cases tested, the estimated map

(Figure 9(b)) is not consistent with the real one. In

the seventh robot pose the estimated and real ones

presented larger difference, leading the filter to

ICINCO 2012 - 9th International Conference on Informatics in Control, Automation and Robotics

412

begin diverge for all tested cases. These experiments

cannot be made using AEA, because it is based in an

incremental way which would lead the robot to

collide with the walls in the environment.

These experiments were important to show that

the SLAMVITA sensors disposition in the panning

turret and its simple types (visual, infrared and

sonar) allow solving the SLAM problem using a

low-cost mobile robot.

The proposed solution presented in this article in

comparison to others works that deal with indoor

SLAM using low-cost mobile robot, such as

presented by (Vazquez and Malcolm, 2005), does

not have restriction to navigate the robot nearby the

environment walls in function of the short

measurement range of the infrared sensors to build

the mapping environment. Moreover, only 4 sensors

are employed in SLAMVITA robot, and due to

theirs different principle of operations, the measure

noises can be solved by the proposed sensor data

fusion algorithm, mainly considering the

complementary fusion method. This is the case were

one of two directional sensor (vision or infrared) has

not measures at a specific angular area of the scan.

6 CONCLUSIONS

The environment mapping is an important task for

many purposes that the mobile robots might

perform, normally requiring sensor data fusion. A

mobile robot was constructed employing three

different sensor types: visual (wireless camera and

laser pointer), infrared (two units) and sonar.

This paper presents the tests of three software

modules that run in a cooperative way to solve the

indoor SLAM problem: a) a sensor data fusion

algorithm; b) a version of particle filter (FastSLAM

1.0); and c) an autonomous featured-based

exploration algorithm.

The results experiments presented show that

there are no significant differences when the

environment exploration task is performed with or

without autonomous exploration. In other word,

choose better positions to acquire the environment

measures are solved by proposed autonomous

algorithm. The estimated maps constructed by the

filter, which represent the environment, are

consistent for robot navigation purpose.

Currently the solution for a larger environment,

around 80 meters in length with some loop situations

is under development with satisfactory partial

results. The experiments in larger environments aim

to consistently validate all software modules

developed in this research. Both the estimated poses

and map must hold the consistency obtained in the

experiments presented in this article.

The main contribution of this research is to solve

the indoor SLAM problem using a low-cost mobile

platform that requires low computational load using

the overall system intelligence running in a PC

computer and embedded in the robot constructed

with simplified hardware and software.

REFERENCES

Aguirre, E., González, A., 2002. Integrating fuzzy

topological maps and fuzzy geometric maps for

behavior-based robots. International Journal of

Intelligent Systems, no. 17, pp. 333–368.

Buonocore, L., Nascimento Jr., C. L., Almeida Neto, A.,

2010. Sistema de baixo custo para medição de

distância em duas dimensões usando câmera IP sem

fio. 18º. Congresso Brasileiro de Automática, Bonito,

MS, Brazil, pp. 2239 – 2246.

Castellanos, J. A., Martínez, J. M., Neira, J., Tardós, J. D.,

1998. Simultaneous map building and localization for

mobile robots: a multisensor fusion approach.

International Conference on Robotics & Automation,

vol. 2, pp. 1244–1249.

Huge, D-W., 2001. Multi sensor data fusion. [e-book]

Available at: < http://www.acfr.usyd.edu.au/pdfs/ trai

ning/multiSensorDataFusion/dataFusionNotes.pdf>

[Acessed 12 September 2009].

Huge, D-W., Bailey, T., 2006. Simultaneous localization

and mapping: Part I. Tutorial. Robotics & Automation

Magazine, vol. 13, no. 3, pp. 99–108.

Ivanjko, E., Petrović, I., Brezak, M., 2009. Experimental

comparison of sonar based occupancy grid mapping

methods. AUTOMATIKA: Journal for Control,

Measurement, Electronics, Computing and

Communications, vol. 50, no. 1-2, pp. 65–79.

Milisavljević, N., Bloch, I., Acheroy, M., 2008. Multi-

sensor data fusion based on belief functions and

possibility theory: close range antipersonnel mine

detection and remote sensing mined area reduction. I-

Tech Education and Publishing, pp. 95–120.

Nguyen, H. G., Blackburn, M. R. (1995). A simple

method for range finding via laser triangulation,

Technical Document 2734, Naval Command, Control

and Ocean Surveillance Center RDT&E Division, San

Diego, CA.

Newman, P., Bosse, M., Leonard, J, 2003. Autonomous

feature-based exploration. International Conference

on Robotics and Automation, vol. 1, pp. 1234–1240.

Pandey, A. K., Krishna, K. M., Nath, M., 2007. Feature

based grid maps for sonar based safe-mapping.

International Joint Conferences on Artificial

Intelligence, pp. 2172–2177.

Thrun, S. B., 1993. Exploration and model building in

mobile robot domains. International Neural Networks,

vol. 1, pp. 175–180.

Solving the Indoor SLAM Problem for a Low-Cost Robot using Sensor Data Fusion and Autonomous Feature-based

Exploration

413

Thrun, S. B., Burgard, W., Fox, D., 2005. Probabilistic

Robotics. Cambridge, USA: MIT Press.

Vazquez, J., Malcolm, C., 2005. Fusion of triangulated

sonar plus infrared sensing for localization and

mapping. International Conference on Control and

Automation, vol. 2, pp. 1097–1102.

Visser, A., 1999. Chapter 9 - Sensor Data Fusion. [pdf]

Available at: < http://www.science.uva.nl/~arnoud/edu

cation/OOAS/fwi/Chap9.pdf> [Acessed 12 September

2009].

Yap Jr., T. N., Shelton, C. R., 2009. SLAM in large indoor

environments with low-cost, noisy and sparse sonars.

IEEE International Conference on Robotics and

Automation, Kobe, Japan, pp. 1395 – 1401.

Zuliani, M., 2009. RANSAC for Dummies-Draft. [pdf]

Available at: < http://vision.ece.ucsb.edu/~zuliani/Res

earch/RANSAC/docs/RANSAC4Dummies.pdf >

[Accessed 15 August 2009].

ICINCO 2012 - 9th International Conference on Informatics in Control, Automation and Robotics

414