This research project developed a global vision system to track the position, angle, direction, linear velocity and angular velocity of three swarm robots. An ID tag design was created using different color patches to identify each robot. The HSV color space was selected to make the algorithm robust against changing lighting conditions. The algorithm uses Euclidean distance to track each robot by calculating the distance between color patches. It determines the angle of each robot and then calculates the linear and angular velocity by comparing values between frames. While the goals of identifying each robot and calculating its parameters were achieved, the processing time per frame was longer than expected.

1. Introduction
This project was done as a part of my curriculum at University of South Australia for the
award of masters of advanced manufacturing technology. The time frame during which this
research was conducted was the final two semesters of my master's degree i.e. the third and
fourth semester (Feb 2010 to Oct 2010. The research project was titled "Development of
global vision system for swarm robots".
Aims and objectives
This research dealt with the use of global vision system to find position, angle, direction,
linear velocity and angular velocity of three robots. This represented a step towards the
development of fully functional swarm robots at university of South Australia.
The expected outcomes were an ID tag which allowed for easy calculation of the orientation
of the robots followed by robust tracking of the individual robots. In addition to this it must
be able to accommodate a large number of robots. Another outcome was identification of an
appropriate colour space which allowed for easy image segmentation separating the regions
of interest and the foreground pixels. Finally the algorithm developed should be such that it
could calculate the required parameters accurately, robustly and consumed a small amount of
processing time per frame.
Methodology
ID tag design
A variety of ID tags have been developed over the years for swarm robots. The ID tag
developed during this research is shown in figure 1.1 below. Here the number 1 represented
the key colour patch. This patch had the same colour for all three robots and coordinates of its
centre gave the position of the robots in the image. The number 0 represented the non key
colour patch. This patch had a different colour for each of the robots and was used for
tracking the individual robots and for angle calculation. The colour 2 represented a colour
which gave 0 and 1 a higher contrast so that these were more easily detectible.

2. Figure 1.1 The ID tag developed
The colour used for the ID tags are shown in figure 1.2
Figure 1.2 Colour used for ID tag
This design of the ID tag represents the simplest patterns used for robot pose estimation. The
reasons for selecting this specific pattern were visual simplicity ease for calculating
orientation and straightforward calculations for tracking individual robots. However a major
flaw in this case was the number of players that can be accommodated. To include more
players would have meant to include more colours for the non key colour patches which
would increase the computational burden for the algorithm significantly.

3. Selection of colour space
Each frame taken showed different pixel values due to the changing lighting conditions.
Therefore a colour space needed to be selected that was impervious and robust against the
changes in lighting conditions. The colour space used in this case was the HSV. This colour
space allows the algorithm to differentiate between colours based on their hue and saturation
values. The main reason for selecting for this colour space is that unlike the RGB colour
space where all the colour planes are correlated, i.e. changes in illumination effects all three
planes, it encodes the colour information in two planes i.e. hue and saturation and the
brightness information in the third value plane. Another method used to minimize the effects
of variations in illumination conditions is normalization of the RGB colour space. The
purpose for doing this is to compare both methods and to find which approach is more suited
for use in swarm robot.
Tracking of individual robots
A very important task was keeping track of the individual robots. Many different measures
had been devised using colour information and specific geometric properties. The tracking
measure used in this thesis belonged to the latter. This measure is termed as the Euclidean
distance. It can be defined as the ordinary distance between two points which one measures
using a common ruler and can be calculated using the Pythagoras theorem. This Euclidean
distance is shown in the figure 1.3
Figure 1.3 Euclidean distance
The formula is given as
𝐸𝑢𝑐𝑙𝑖𝑑𝑒𝑎𝑛 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒 = √ (𝑦2 − 𝑦1 )2
+ (𝑥2 − 𝑥1)2
Euclidean distance
X2-X1
Y2-Y1

4. Using the formula defined above the algorithm will calculate the Euclidean distance between
one of the non key colour patches and all of the three key colour patches and the key colour
patch having the lowest Euclidean distance will be selected. This procedure will be repeated
for all the three non key colour patches. This makes use of the fact that the distance between
the two patches is always less than 40 pixels for all the three patterns. Using this approach
allows the algorithm to track the positions and angles of the three patterns which in turn
makes it possible to find the linear velocity and the angular velocity.
The algorithm
A frame used in the experiment is shown in figure 1.4
Figure 1.4 A picture used in the experiment
The centre circle in this case is of red colour whereas the non key colour patch circles are
magenta, yellow and green. The algorithm starts by acquiring the RGB frame. As it is evident
from figure 1.4 there are three colour IDs. For the sake of convenience these are called
magenta robot, yellow robot and green robot. The acquired image is then converted to HSV.

5. This is followed by applying a threshold and an opening separating the centre red circles
shown in figure 1.4. It should be noted that it is not known at this point that which centre
circle belongs to which robot. The algorithm will for now store the x and y coordinates of the
three centre circles. During tracking the algorithm is able to differentiate between these based
on the Euclidean distance which in this case is the distance between the two circles on a robot
e.g. magenta circle on magenta robot and its respective centre. The next step is applying
multiple thresholds to find centre coordinates for magenta, yellow and green circles. The
algorithm then finds the Euclidean distance for each robot and is able to track these
efficiently. The algorithm has now differentiated between the three robots and calculates the
angle for each robot next. The angle is found by calculating the slope of the line connecting
the centres of the key colour patch and non key colour patch for each robot e.g. for magenta
robot the line is between centre of the red circle and centre of magenta circle. This is
followed by finding the velocity both linear and angular. To find these the positions and
angles obtained for each frame are stored separately. To find the linear velocity and angular
velocity of a robot between two frames a simple difference is taken between the centre
coordinates and angle values obtained for both frames divided by the time between two
frames. Finally based on robot angle the direction is calculated. For example if angle is 90
degrees the robot is headed in the north direction and if it is 270 degrees the robot is going in
the south direction.
The other approach taken was normalizing the RGB colour planes to find the required
parameters. The procedure for both approaches is the same and the only difference is the
colour space used. Both of these approaches were compared in terms of processing time, post
processing required and detecting secondary colours. It was concluded that the algorithm
employing HSV colour space was better of the two approaches.
Results
The project was able to meet all the expected outcomes. However the processing time per
frame was more than expected. Currently attempts are being done to reduce this time. The
vision algorithm developed by me will be combined with a PHD student's obstacle avoidance
algorithm to improve its overall efficiency and robustness.