Color Object Detection
Color Object Detection
Color Object Detection
I. INTRODUCTION
In this paper we present two new techniques for the
problem of colored object detection by autonomous robots
operating in dynamic, adversarial environments under variable
lighting conditions. Concretely, we examine the problem of
colored object detection in an outdoor environment by an
autonomous robot using a monocular color camera. We
investigate this problem using the new Segway RMP platform
for operating in the task of Segway Soccer, where teams of
Segway riding humans and robots compete in a soccer-like
game outdoors. Hence, the environment is adversarial, and
highly dynamic, the combination of which presents many
interesting and unique challenges to perception.
The combination of variable lighting conditions, dynamic
environments, and autonomous robots makes object
segmentation and recognition challenging due to the competing
needs for fast, computationally tractable algorithms that are at
the same time able to adapt or overcome changes in lighting.
Here, fast means that the algorithm runs as close as possible to
the camera's full frame rate of 30Hz, while still leaving
processing resources available for robot autonomy algorithms.
Such speed requirements are necessary when the domain is
adversarial, and robots are able to achieve speeds of 3.5m/s
(8mph) and operate in close proximity to one another.
This research was sponsored by the United States Army under Grant No.
DABT63-99-1-0013. The content of the information in this publication does
not necessarily reflect the position or the policy of the Defense Advanced
Research Projects Agency (DARPA), the US Army or the US Government,
and no official endorsement should be inferred.
III.
(1)
uj
vj)
(2)
a+ b + c
Pixel Frequency
Snapshot of Histogram
52
103
154
205
256
Histogram Index
Figure 3. A sample image and the resulting peak detected from the
histogram of the vector projection image.
H ( m) min H (i ), m < p
i =m
1
h( i )
w x w 2 i < x + w 2
(5)
NH
j =i
(6)
H ( j)
i { N }, = y, u, v
1
(7)
2.
NH
b c ) (y j
i= p
(a
(3)
Sj =
1.
y
Dy
u
Du
v
Dv
(8)
=k
vi v
Dv
Du
Dy
ui u
Si =
yi y
(9)
Figure 4. The left image shows the raw output from the camera in a lab
setting. The right image shows the results after processing. Ball pixels are
shown as orange (shown in gray here), a bounding box is drawn aroun the
resulting region, and its centroid is marked with a +.
A. Performance
Object detection accuracy was measured by running the
three vision algorithms on three image sequences focused on an
orange, size 4, soccer ball. All sequences were recorded from
the Segway RMP platform and run through the different vision
systems separately. The high-level filters were identical for
each approach. The color table for the complete CMVision
approach was created by hand using sample images from each
domain. The histogram parameters for the remaining two
techniques were identical and were found by hand calibration.
The mean and distance estimates for the ellipsoid approach
HDET, were found from sample images in each setting. Figure
5 shows a sample snap-shot from each testing domain. The
three test sequences were:
1.
2.
3.
% Frames in which
Object is Undetected
A
B
C
63.5
47.0
57.7
11.2
27.7
16.9
19.2
0.0
0.0
6.2
25.4
25.4
TABLE II.
CONFUSION MATRIX USING FIXED COLOR THRESHOLDS (A),
HTVP (B), HTED (C) USING INDOOR IMAGE SEQUENCE WITH VARIABLE
LIGHTING
% Frames in which
Object is Detected
A
B
C
100
100
46.6
X Coordinate
53.4
240
230
220
210
200
190
180
170
160
150
Observed
Actual
Time
X Coordinate
Object
Present
in Image
Object
Absent
in Image
% Frames in which
Object is Undetected
A
B
C
For the outdoor sequence, where both robot and ball were
in motion, we show the reported x image coordinate as a
function of time. Also shown in each graph is the ground truth
as labeled by a human operator.
240
230
220
210
200
190
180
170
160
150
Observed
Actual
Time
140
80
60
40
Luminence Value
Fixed Co lor Threshold
HTVP
HTED
20
0
X Coordinate
100
Detection Flags
Average Y Value
120
240
230
220
210
200
190
180
170
160
150
Observed
Actual
Time
Time / ms
Vector Projection
Ellipsoid Distance