VISION BASED OBSTACLE AVOIDANCE AND ODOMETERY

FOR SWARMS OF SMALL SIZE ROBOTS

M. Shuja Ahmed, Reza Saatchi and Fabio Caparrelli

Material and Engineering Research Institute, Shefﬁeld Hallam University, Shefﬁeld, U.K.

Keywords:

Obstacle Avoidance, Visual Odometery, Swarm Robotics.

Abstract:

In multi-robotic systems, an approach to the coordination of multiple robots with each other is called swarm

robotics. In swarm robotic systems, small size robots with limited memory and processing resources are

used. Integration of vision sensors in such robots can complicate the design of the robots but at the same

time, a single vision sensor can be used for multiple objectives as it provide rich surrounding information.

As the vision algorithms are normally computationally demanding and robots in swarm systems has limited

memory and processing capabilities, so the requirements of light weight vision algorithms also arises. In

this research, the use of vision sensor information is made for achieving multiple objectives. A solution

to obstacle avoidance, which is the basic requirement as robots move in a cluttered environment and also

odometry which is essential for robot localization, is provided using only visual clues. The approach developed

in this research is computationally less expensive and suitable for small size robots, where processing and

memory constraints limit the use of computationally expensive approaches. To achieve this a library of vision

algorithms is developed and customized for Blackﬁn processor based robotic systems.

1 INTRODUCTION

Current research in the ﬁeld of swarm robotics is

largely focused on providing a computer vision so-

lution to guide multiple robots. In most cases small

size robots with limited processing and memory re-

sources, are used. Integration of vision sensor tech-

nology in such systems helps to increase the onboard

intelligence but at a cost of a signiﬁcant increase in

computational load. The fact that computer vision so-

lutions are computationally very heavy for small size

robots, makes collective achievement of tasks chal-

lenging among swarm of robots. In this research, a

swarm robotic system is considered (shown in ﬁg-

ure 1) in which multiple robotic organisms have the

ability to physically dock, share information, energy

and computational resources with each other and also

can locate the charging points to charge their batter-

ies (Kernbach et al., 2010). These swarms of robots

can self assemble and can artiﬁcially create a single

three dimensional robotic organism, which requires

these robots to interact with the environment and also

with each other. These swarms of robots are ex-

pected to perform many different operations in paral-

lel (e.g. docking, communication, scene understand-

ing). Hence, providing vision based senses for guid-

ance and decision making in such robots, naturally re-

quires the development of an efﬁcient and optimized

library of vision processing algorithms. This, in turn,

helps to achieve real time performance on small size

robots. The fact that a vision sensor provides a rich

surrounding information (i.e. a single sensor infor-

mation can be used for multiple purposes), therefore

in this research, this visual information is utilized to

perform multiple tasks that is, obstacle avoidance and

odometery which essentially will help to guide and

localize swarm of robots in a structured environment.

Figure 1: Swarm of robots working collectively (Kernbach

et al., 2010).

For autonomous mobile robotic systems, com-

puter vision provides many solutions to obstacle

avoidance but when it comes to small size robots,

which are also expected to perform many other oper-

115

Shuja Ahmed M., Saatchi R. and Caparrelli F..

VISION BASED OBSTACLE AVOIDANCE AND ODOMETERY FOR SWARMS OF SMALL SIZE ROBOTS.

DOI: 10.5220/0003820701150122

In Proceedings of the 2nd International Conference on Pervasive Embedded Computing and Communication Systems (PECCS-2012), pages 115-122

ISBN: 978-989-8565-00-6

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

ations in parallel, then the number of obstacle avoid-

ance approaches in hand is limited. In (Michels

et al., 2005) an obstacle avoidance technique using a

monocular vision camera together with a laser range

ﬁnder is addressed. The testing of the algorithm is

performed in a outdoors highly unstructured environ-

ment but the testing system used is not strictly an em-

bedded system as all the image processing is done on

a development platform so the chances of techniques

not meeting the real time constraints were also possi-

ble. Another common approach, addressed in (Chao

et al., 1999) (Borenstein and Koren, 1985) (Boren-

stein and Koren, 1988), is based on edge detection.

In this method, the algorithm determines the vertical

edges of the obstacle and helps robots to move around

the edges without colliding against the obstacles. In

(Pratt, 2007), the Lucas-Kanade optical ﬂow based al-

gorithm is used for MAVs (Micro Aerial Vehicles)

in urban environment. Similarly, in (Souhila and

Karim., 2007) Horn and Schunck’s optical ﬂow based

algorithm is used in autonomous robots. Using opti-

cal ﬂow, image velocity vectors are determined which

can be split into translational and rotational compo-

nents. From the translational component, the time

to contact information for the obstacle can be calcu-

lated which helps in taking necessary actions. An-

other area of research in computer vision and robotics

targeted in this study is the development of efﬁcient

visual odometery algorithm for small size robots. In

(Maimone et al., 2007), a feature tracking based mo-

tion estimation approach is presented to obtain visual

odometery information using stereo images captured

from NASA’s Mars Exploration Rovers (MERs) in a

highly unstructured environment. In (Campbell et al.,

2005), visual odometry results using optical ﬂow in-

formation are presented when the ground robot is

moved on a varying terrain including indoor and out-

door environments. The errors reported were 3.3%

and 7.1% when the robot is moved on a carpet (high

friction) and polished concrete, respectively. Simi-

larly, in (Milford and Wyeth, 2008) (Kyprou, 2009) a

scanline intensity based simple algorithm to obtain vi-

sual odometry is presented. The odometry error with

this algorithm can be large but, in spite of this, no-

table results are achieved when the odometry infor-

mation is used with a SLAM (Simultaneous Localiza-

tion and Mapping) system. A relevant research done

in (Schaerer, 2006) addresses the use of line features

tracking. Using Hough Transform lines are tracked

for obtaining the distance and orientation information.

Another algorithm presented in (Younse and Burks,

2007) is based on feature tracking using the Lucas-

Kanade algorithm. It also utilizes the information ob-

tained from camera modelling (intrinsic and extrin-

sic parameters) to precisely locate the new position

of the vehicle. The average translation error reported

when the vehicle moved 30 cm is 4.8 cm and the aver-

age rotation errors were 1 and 8 degrees for a 45 and

180 degrees rotation, respectively. This approach per-

formed poorly at high rotation rate as features could

move out of the search window. Hence, in the ﬁeld

of autonomous robotics, many approaches to visual

odometery (Maimone et al., 2007) (Nistar et al., 2006)

(Howard, 2008) using high speed systems are ad-

dressed and notable results are achieved. The high

computational cost of these algorithms makes them

unsuitable for swarms of small size robots. However,

further advances in research to provide fast and reac-

tive solutions to these problems is still required. In

the following sections, the methods used to perform

vision based obstacle avoidance and odometery are

detailed in section 2. In section 3, the results obtained

when the experiments performed in the in-door en-

vironment, are presented. Conclusions are drawn in

section 4.

2 METHODOLOGY

The hardware is an important factor which strongly

inﬂuences the methods adapted to solve the problem

at hand. The onboard processing on the robot (shown

in Figure 1) was achieved using a high performance

16/32-bit Blackﬁn BF537E processor. uClinux (mi-

cro controller linux), which is a powerful operating

system customized for embedded systems, was used

as the onboard operating system. Code compilation

was done using GNU cross compilers on a Linux

based development platform. For the testing and

demonstration of the developed vision algorithms,

SRV robot by Surveyor Corporation was used. Before

proceeding to the complex vision algorithms such as

obstacle avoidance and visual odometry, a library of

basic vision algorithms was developed. This library

was optimised especially for the Blackﬁn processor

architecture. It includes image conversion to differ-

ent formats (such as YUV to colour image), colour

to greyscale image, image gradient using Sobel and

Canny operator, region growing based image segmen-

tation, colour blob detection, feature detection using

Harris algorithm, cross-correlation based algorithm to

solve the feature correspondence problem, image ero-

sion and dilation algorithms. It was decided to add

more algorithms as the need arises. In the following

sections, the approaches used to perform vision based

obstacle avoidance and range of visual odometry al-

gorithms developed are discussed in detail.

PECCS 2012 - International Conference on Pervasive and Embedded Computing and Communication Systems

116

2.1 Obstacle Avoidance

There are many ways to accomplish obstacle avoid-

ance but efforts are being made to come up with an al-

gorithm with less computational complexity. For ex-

ample, algorithms with more ﬂoating point operations

have more computational complexity and execution

time. The Blackﬁn processor used here is a ﬁxed point

processor and can not perform ﬂoating point opera-

tions efﬁciently so algorithms with less ﬂoating point

operations are preferred. Finally, the algorithms are

customized to exploit the ﬁxed point Blackﬁn archi-

tecture as suggested in (Lukasiak et al., 2005). Two

obstacle avoidance algorithms, where one was based

on segmentation and the other utilized the optical ﬂow

information, are presented.

2.1.1 Segmentation based Obstacle Avoidance

To explain the concept of segmentation based obstacle

avoidance, we considered the image in Figure 2 show-

ing a couple of obstacles. Some assumptions were

made that the robot is placed on a ﬂat ground with

camera slightly tilted down. In Figure 2b, segmenta-

tion result is shown. To determine ground region, one

assumption can be that the biggest segmented region

is from ground but it may be false when the robot is in

front of a big obstacle which result the biggest region.

In the current implementation, the speed of the robot

is set to guarantee that it is not very close to the obsta-

cle and the region covering the middle bottom of the

image is considered to be the ground region. Some-

times, while turning, there is a possibility that the ob-

stacle is very close to the robot. In this scenario, the

robot can collide with it. To overcome this problem,

the robot keeps track of the intensity of ground region

in the last few frames. A sudden change in the inten-

sity helps robot to determines the presence of an ob-

stacle. Following this approach, the isolated ground

region is shown in Figure 2c (Total Ground Map).

To determine the ground map visible to the robot, the

’Total Ground Map’ is ﬁlled vertically up with white

pixels until the obstacle boundary is detected and the

rest of the pixels to the top are ﬁlled with black pix-

els. The same is done for all the columns. The im-

age is ﬁnally dilated for reﬁnement. The ﬁnal visible

ground map is shown in Figure 2d. To determine the

presence of an obstacle, if from the centre bottom of

the ﬁnal ground map, the number of white pixels are

greater than 30 in the vertical direction, then the robot

goes straight. It also checks for enough space from

left and right direction. If white pixels are less than

30, then the algorithm checks from which side it has

more clearance and turn the robot in that direction. In

Figure 2d an expected robot trajectory is shown.

Figure 2: The step by step processing of segmentation based

obstacle avoidance algorithm.

2.1.2 Optical Flow based Obstacle Avoidance

Another approach to obstacle avoidance is based on

Horn and Schunck (Horn et al., 1993) optical ﬂow

information and is inspired from the work done in

(Souhila and Karim., 2007). Horn and Schunck de-

rived equations that relate image brightness at a point

to the motion of brightness pattern. For this purpose,

partial derivatives of the image brightness in x (E

), y

) directions and in time space (E

) are obtained.

These partial derivatives are used to obtain optical

ﬂow vectors u and v in x and y direction, respectively.

These vectors are obtained using the following recur-

sive equations.

n+1

= ¯u

−

¯u

+ E

¯v

+ E

]

+ E

(1)

n+1

= ¯v

−

¯u

+ E

¯v

+ E

]

+ E

(2)

Where α is a weighting factor and ¯u

and ¯v

are local

averages. These recursive equations performs many

ﬂoating point operations and it was possible to pro-

cess only one frame in 2 seconds. For faster exe-

cution, part of algorithm was coded to perform dec-

imal operations. For this purpose, the necessary pre-

cision was determined. For example, for a number

1.3678987, to keep precision upto four decimal point,

the number will be multiplied with 10

and rest of

computation will be done in decimal point format.

Where decimal format did not work, ﬁxed point im-

plementation was adopted. This made it possible to

process 3.5 frames per second. Now to determine

the presence of the obstacle, the image was divided

into left and right parts and every time the magni-

tudes of the optical ﬂow vectors were calculated for

these parts. If the sum of these magnitudes exceeds

some predeﬁned thresholds, it was assumed that the

obstacle is in front of the robot and it turns in the di-

VISION BASED OBSTACLE AVOIDANCE AND ODOMETERY FOR SWARMS OF SMALL SIZE ROBOTS

117

rection which produces less magnitude of the optical

ﬂow vectors (i.e. more clearance).

2.2 Visual Odometry

Different approaches to perform visual odometry

were implemented and compared with each other. Ini-

tially an algorithm, using scan line intensity proﬁle

formed by sub images, for obtaining the visual odom-

etry information, was adapted. This scanline proﬁle

idea was also adapted in (Pomerleau, 1997) for devel-

oping visual steering system and it can also be found

in (Milford and Wyeth, 2008) where visual odometry

based on single camera was used for SLAM system.

The scan line proﬁle is a one-dimensional vector ob-

tained by summing the intensity values in each pixel

column of the selected part of the image and then nor-

malized. This vector proﬁle is compared with the

other proﬁles obtained from the consecutive images

to extract the rotation and forward distance covered

information. Figure 3 shows the top and bottom win-

dow images which are used to determine rotation and

forward speed, respectively. The scanline idea for ob-

taining the speed was found very sensitive to lighting

conditions and was discarded. So, to obtain the dis-

tance covered information, two approaches were ex-

amined, i.e. the feature tracking using Lucas-Kanade

(Lucas and Kanade, 1981) and feature matching us-

ing Normalized Cross Correlation (Lewis, 1995) al-

gorithm. These are explained below.

Figure 3: Image parts for rotation and translation cues.

2.2.1 Rotation Estimation

Rotation information is obtained by comparing the

scan line intensity proﬁle vectors from two consec-

utive images captured by the vision system. In Fig-

ure 4b, the scan line intensity proﬁles from two con-

secutive images are shown. For comparison between

the proﬁles, the average absolute intensity difference

f (t) was calculated between the two scanline vectors

while a relative shift of the vectors with respect to

each other is also performed. The following equation

is used to achieve this task.

f (t, V

, V

) =

w − |t|

(

w−|t|

∑

n=1

n+max(t,0)

−V

n−min(t,0)

(3)

where V

and V

are the intensity proﬁle vectors to

be compared (shown in red and blue colours, respec-

tively, in Figure 4b), t is the amount of pixel shift per-

formed between the two proﬁles, and w is the width

of the windowed image. In Figure 4a, the absolute in-

tensity differences graph, when one proﬁle vector is

shifted over the other, is shown. The pixels shift p

were obtained by determining the value t which min-

imizes the function f (t) for the vectors V

and V

= min

t∈|ρ−w,w−ρ|

f (t, V

, V

) (4)

The selection of offset ρ was made such that there

was enough overlap between the two proﬁles. The

values used for w and ρ were 88 and 40, respectively.

For the two proﬁles shown in Figure 4b , the value of

pixel shift p

obtained is 15 pixels in the left direction

from the center of image. This pixel shift is multiplied

by gain factor α to perform conversion into an angular

shift value ∆θ. The value of α can be obtained empir-

ically or by using the camera’s intrinsic parameters:

∆θ = αp

(5)

Figure 4: a) Average intensity difference when one proﬁle

shifted over the second. b) Two scanline proﬁles.

PECCS 2012 - International Conference on Pervasive and Embedded Computing and Communication Systems

118

2.2.2 Translation Estimation using

Lucas-Kanade Feature Tracking

The translation estimation algorithm is divided into

two tasks. The ﬁrst is to detect salient features in the

image and the second is the features tracking. For

feature detection, Harris feature detection (Harris and

Stephens, 1988) algorithm was implemented. The

features detected were then tracked in the following

images by the Lucas-Kanade tracking algorithm. The

basic idea of the Lucas-Kanade (Lucas and Kanade,

1981) tracking algorithm is the following. Let, I

and

be two images and u be a feature detected on im-

age I

, where u = [x, y] and I

(u) =I

(x, y). Then, the

goal of the Lucas-Kanade tracking is to ﬁnd the loca-

tion v = u + d on the second image I

such that I

(u)

and I

(v) are similar. Let, d be the image velocity

vector at point u. The next step is to deﬁne a neigh-

bourhood around feature u (i.e. w

and w

in x and y

directions), where similarity analysis between u and v

can be made. Then, vector d can be deﬁned as the one

which minimizes the following residual function:

∈ (d) =

∑

x=u

−w

∑

y=u

−w

(x, y) − I

(x + d

, y + dy)]

(6)

After obtaining the velocity vector, the new fea-

ture location can be determined and tracked. For

odometry, the average shift made by the feature in y

direction (after multiplying by scale factor) is used to

determine the translation and the shift in x direction is

used for the small rotation information. To measure

large rotation, scan line algorithm is used in parallel.

2.2.3 Translation Estimation using Normalised

Cross Correlation Feature Matching

Another approach developed in this work was based

on Normalised Cross Correlation matching of salient

features. The features are detected in two images us-

ing the Harris algorithm. Matching and correspon-

dence between the features is performed using nor-

malised cross correlation algorithm. In this algorithm,

to perform feature matching, a small template is taken

around each feature in one image and is matched with

the other image within a predeﬁned search window.

The basic equation used to perform cross correlation

matching is given below.

C(m, n) =

∑

x,y

bI(x, y) −

m,n

c[t(x − m, y − n) −

[

∑

x,y

[I(x, y) −

m,n

]

∑

x,y

[t(x − m, y − n) −

]

0.5

(7)

In this equation, I(x,y) is the image and t(x,y) is the

target template (i.e. template around feature),

t is the

mean of the template and

m, n

is the mean of the im-

age. Once corresponding features are identiﬁed, then

the average pixel shift in x and y direction is used

to determine the rotation and translation, respectively.

For large rotation, scan line works in parallel.

3 RESULTS

This section is dedicated to the results. It includes

obstacle avoidance using segmentation/optical ﬂow

information and visual odometery results from scan

line based rotation, Lucas-Kanade tracking and nor-

malised cross correlation based translation algorithm.

3.1 Obstacle Avoidance

To test this algorithm, a test platform was devel-

oped. Obstacles were placed on uniform white sur-

face. Some processed images are shown in Figure 5.

When the algorithm ﬁnds that the white pixels to the

obstacle boundary are less than 30, then it determines

which side gives more clearance and turns in that di-

rection. Factor 30 pixels is determined empirically.

Figure 5: Segmentation based obstacle avoidance.

Using this algorithm, many tests were performed.

One of the path followed by the robot is shown in

Figure 6a. The maximum frame rate achieved with

this approach is 2.46 frames per second.

Figure 6: (a) Segmentation result (b) Optical ﬂow result.

A trajectory made by the robot using optical ﬂow

VISION BASED OBSTACLE AVOIDANCE AND ODOMETERY FOR SWARMS OF SMALL SIZE ROBOTS

119

algorithm, is shown in Figure 6b. In Figure 7, the op-

tical ﬂow information generated from the two consec-

utive images, when the robot was moving in forward

direction, is shown. A zoomed-in version of the opti-

cal ﬂow ﬁeld is also shown. The optical ﬂow vectors

obtained are in the correct direction (i.e. opposite to

the direction of motion) but there are some vectors in

random directions and act as noise. The maximum

frame rate achieved with this approach is 3.5 frames

per second.

Figure 7: Optical ﬂow ﬁeld between 2 consecutive images.

3.2 Visual Odometry

For visual odometry, the scanline algorithm was used

for getting the rotation estimates. In Figure 8, the re-

sults obtained, when the robot was making a square

trajectory, are shown. The robot moved equal dis-

tances after every 90 degrees turn. In Figure 8a, the

angle measurement obtained from the visual odome-

try is shown. Robot has taken 89.5, 92, 91 and 88 de-

grees turns in place of 90 degrees turn. The resultant

square trajectory is also shown in Figure 8b. This ro-

tation and translation measurements are obtained us-

ing the scanline proﬁle based algorithm. The scanline

algorithm for translation is found very sensitive to the

lighting conditions. It gives totally different transla-

tion estimates in a different environment. So here, ac-

curacy achieved in rotation using scanline is focused.

To obtain the accuracy in translation estimates, the

feature based approach was used. For features de-

tection, black marks were made on the ground sur-

face. In Figure 9, some of the results obtained from

the Lucas-Kanade tracking (left column) and feature

matching results from the normalised cross correla-

tion algorithm (right column) are shown.

To test feature based translation estimation algo-

rithm, two tests were performed. In Figure 10a, the

visual odometry results obtained from test 1, when

the robot was moved 36 inches forward, took a 90 de-

gree right turn and then moved 11 inches, are shown.

In the results obtained from the Lucas-Kanade algo-

rithm, the robot moved 38.6 inches forward and 10

Figure 8: Angle information and the actual trajectory ob-

tained from visual odometry for square trajectory.

Figure 9: Lucas-Kanade and Normalised cross correlation

feature tracking and matching results.

inches after taking a right turn. According to the mea-

surements obtained using the cross correlation based

algorithm, the robot moved 38.7 inches forward and

then 9.5 inches after taking a right turn. Both of these

measurements are very close to the actual trajectory

made by the robot. Similarly, Figure 10b shows the

results obtained from test 2 when the robot moved 36

inches forward, took a 90 degree left turn and then

moved 9.5 inches. According to the measurements

obtained using the Lucas-Kanade algorithm, the robot

moved 37.4 inches forward and 8.3 inches after taking

a left turn. According to the cross correlation based

PECCS 2012 - International Conference on Pervasive and Embedded Computing and Communication Systems

120

Figure 10: Visual odometry (a) test 1 (b) test 2.

algorithm, the robot moved 39.4 inches forward and

8.6 inches after taking a left turn.

4 CONCLUSIONS

In this research, vision algorithms for a Blackﬁn

based small size robotic system have been developed.

For obstacle avoidance, a slightly slower frame rate

was achieved but it can be improved by implementing

the computationally expensive sections of the code in

assembly language. For visual odometry, the Lucas-

Kanade based approach has provided good results

when tested using Matlab. But it is found that, in case

of features lost, some time it provides less accuracy

than expected. It is concluded that, the use of scan

line based rotation information with the feature based

approach provide an efﬁcient solution to visual odom-

etry but at the same time, it can be easily deceived if

the robot come close to a moving object such as an-

other robot. To overcome this, information from vi-

sual odometry may be fused together with the wheel

odometry to determine whether the robot was actually

moved or object in the environment moved.

ACKNOWLEDGEMENTS

This research was funded by European Commission

Seventh Framework Programme FP7/2007-2013 re-

search project REPLICATOR.

REFERENCES

Borenstein, J. and Koren, Y. (1985). A mobile platform for

nursing robots. In IEEE Transactions on Industrial

Electronics, Vol. 32, No. 2.

Borenstein, J. and Koren, Y. (1988). Obstacle avoidance

with ultrasonic sensors. In IEEE journal of robotics

and automation, vol. ra-4, no. 2.

Campbell, J., Sukthankar, R., Nourbakhsh, I., and Pahwa,

A. (2005). A robust visual odometry and precipice

detection system using consumer-grade monocular vi-

sion. In Proc. ICRA, Barcelona, Spain.

Chao, M., Braunl, T., and Zaknich, A. (1999). Visually-

guided obstacle avoidance. In 6th International Con-

ference on Neural Information Processing ICONIP.

Harris, C. and Stephens, M. (1988). A combined corner and

edge detector. In Proceedings of the 4th ALVEY vision

conference, University of Manchester, England.

Horn, Berthold, K. P., and Schunck, B. G. (1993). Deter-

mining optical ﬂow. In Artiﬁcial Intelligence: 81-87.

Howard, A. (2008). Real-time stereo visual odometry for

autonomous ground vehicles. In IEEE/RSJ Interna-

tional Conference on Intelligent Robots and Systems.

Kernbach, S., Scholz, O., Harada, K., Popesku, S., Liedke,

J., Raja, H., Liu, W., Caparrelli, F., Jemai, J., Hav-

lik, J., Meister, E., and Levi, P. (2010). Multi-robot

organisms: State of the art. In ICRA 2010 Workshop

Modular Robots: State of the Art.

Kyprou, S. (2009). Simple but effective personal localisa-

tion using computer vision. In MEng Report. Depart-

ment of Computing, Imperial College London.

Lewis, J. P. (1995). Fast template matching. In Vision In-

terface, p. 120-123.

Lucas, B. D. and Kanade, T. (1981). An iterative image

registration technique with an application to stereo vi-

sion. In IJCAI.

Lukasiak, R., Katz, D., and Lukasiak, T. (2005). Enhance

processor performance in open-source applications. In

Analog Dialogue 39-02.

Maimone, M., Cheng, Y., and Matthies, L. (2007). Two

years of visual odometry on the mars exploration

rovers. In Journal of Field Robotics. Special Issue:

Special Issue on Space Robotics. Volume 24, Issue 3.

Michels, J., Saxena, A., and Andrew, Y. (2005). High speed

obstacle avoidance using monocular vision and rein-

forcement learning. In Proceedings of the 22nd Inter-

national Conference on Machine Learning.

Milford, M. and Wyeth, G. (2008). Mapping a suburb with

a single camera using a biologically inspired slam sys-

tem. In Robotics, IEEE Transactions on , vol.24, no.5,

pp.1038-1053, Oct. 2008.

Nistar, D., Naroditsky, O., and Bergen, J. (2006). Visual

odometry for ground vehicle applications. In Journal

of Field Robotics, Volume 23.

Pomerleau, D. (1997). Visibility estimation from a mov-

ing vehicle using the ralph vision system. In In IEEE

Conf. Intelligent Transportation Systems.

Pratt, K. (2007). Smart sensors for optic ﬂow, obstacle

avoidance for mavs in urban environments. In Inter-

national Journal of Advanced Robotic Systems.

VISION BASED OBSTACLE AVOIDANCE AND ODOMETERY FOR SWARMS OF SMALL SIZE ROBOTS

121

Schaerer, S. S. (2006). Practical visual odometry for small

embedded systems. In Master’s Thesis, Department

of Electrical and Computer Engineering, University

of Manitoba.

Souhila, K. and Karim., A. (2007). Optical ﬂow based robot

obstacle avoidance. In International Journal of Ad-

vanced Robotic Systems.

Younse, P. J. and Burks, T. F. (2007). Greenhouse robot

navigation using klt feature tracking for visual odom-

etry. In International Commission of Agricultural En-

gineering. CIGR E-Journal Volume 9.

PECCS 2012 - International Conference on Pervasive and Embedded Computing and Communication Systems

122