The document describes a method for detecting balls in images and videos taken during a soccer game. It discusses challenges in ball detection due to the ball's small size and similarity to other objects. The method first finds ball candidates using color segmentation in YUV color space and grouping connected pixels. It then classifies candidates as balls or non-balls using a support vector machine trained on color and texture features. Experiments show the method can reliably identify balls in soccer footage.
1. A B I N J O: U C
L C R I
Hannes Schulz, Hauke Strasdat, and Sven Behnke
University of Freiburg
Institute of Computer Science
Nov 29, 2007
2. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
M
The ball
is small,
3. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
M
The ball
is small,
is easy to confuse with other objects
4. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
M
The ball
is small,
is easy to confuse with other objects
is the most important object on the field:
You cannot play sensibly without knowing its position
5. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
M
The ball
is small,
is easy to confuse with other objects
is the most important object on the field:
You cannot play sensibly without knowing its position
We should put a lot of effort into finding the single real ball.
21. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
C E YUV S
22. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
C E YUV S
Two ellipses for “orange”
Actual ball color
Wider, brownish color
23. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
C E YUV S
Two ellipses for “orange”
Actual ball color
Wider, brownish color
Allows for motion blur
24. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
S C I
orange
YUV Camera Image
(-candidate)
white
green
64 : 1
25. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
S C I
1. Find Maximum
2. Find Weighted Mean orange
(-candidate)
white
green
64 : 1
26. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
S C I
orange
cut corresponding area
(-candidate)
white
green
64 : 1
27. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
S C I
orange
cut corresponding area
(-candidate)
white
Box size depends on position in image
green
64 : 1
29. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
P C
30. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
P C
Projection changes with ellipses: Robust to changes in
lighting conditions
31. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
P C
Projection changes with ellipses: Robust to changes in
lighting conditions
32. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
P L
YUV-Image Y-Image Subsampled Y-Image
Subtraction of mean: Robust to changes in lighting conditions
33. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
H L F
34. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
H L F
38. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
NR KS 06 – T
H
520 MHz ARM PocketPC
VGA HTC Camera
D S
160 balls
440 non-balls
divided randomly in training set
(80%) and test set (20%)
39. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
NR KS 06 – T
H
520 MHz ARM PocketPC
VGA HTC Camera
D S
160 balls
440 non-balls
divided randomly in training set
(80%) and test set (20%)
P (T S)
100% of distractors classified
correctly
1 ball out of 32 not recognized
40. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
NR KS 06 – R T
H
520 MHz ARM PocketPC
VGA HTC Camera
R T T
Robot decides autonomously
which object to approach.
41. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
NR KS 2007
H
1.33 GHz PC
WVGA µEye Camera
42. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
NR KS 2007
S D D S
273 balls
548 non-balls
training set 62%, validation set
13%, test set 25%
varying lighting conditions
43. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
NR KS 2007
A S D S
Avg Luminance 273 balls
ball non-ball 548 non-balls
training set 62%, validation set
13%, test set 25%
varying lighting conditions
Avg Orange-Greenness
ball non-ball
44. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
NR KS 2007
A S D S
Avg Luminance 273 balls
ball non-ball 548 non-balls
training set 62%, validation set
13%, test set 25%
varying lighting conditions
Avg Orange-Greenness R T S
ball non-ball 91.1% accuracy
76.6% if stimuli flipped up/down.
Drop suggests dependency on
gradient.
88.2% if lighting in testset differs.
Classifier seems indifferent.
45. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
C C
Classifier
Task Neural Net Linear Classifier KNN (k = 5)
Regular 91.1% 86.5% 88.5%
Flipped 74.0% 70.0% 76.6%
46. Balls and Non-Balls Finding Ball Candidates Classification of Ball Candidates Experiments
C
The ball has properties aside from being orange
These properties are exploited by our Neural Network
Classifiers
Changes in Lighting conditions can be dealt with by projection
to lines in YUV-space and Luminance.
The method introduced here can be generalized to other
small-sized objects on the field.