1. by Utku Tatlıdede Kemal Kaplan
AIBO VISION by
Utku Tatlıdede
Kemal Kaplan

2. AIBO ROBOT Specifications: ERS-210 CPU clock speed of 192MHZ
20 degrees of freedom
Temperature,Infrared Distance, Acceleration, Pressure, Vibration Sensors
CMOS Image Sensor
Miniature Microphones, Miniature Speaker, LCD Display
Dimensions and Weight:
Size (WxHxL) 6.06" (W) x 10.47" (H) x 10.79" (L)
Weight 3.3 lbs. (1.5kg)

3. AIBO VISION CCD camera 16.8 million colors (24 bit)
176x144 pixel image
Field of view 57.6° wide and 47.8° tall
Up to 25 fps
Stores information in the YUV color space

4. PROCESSING OUTLINE Color Segmentation -Representations -Algorithms
Region Building and Merging - Region Growing - Edge detection
Object Recognition - Classification - Template matching - Sanity check
Bounding Box Creation

5. COLOR REPRESENTATION CIE-XYZ RGB CMY, CMYK HSV, HSI, HLS YIQ
Color can be physically described as a spectrum which is the intensity of light at each wavelength.
Radiance: Energy (W) from light source
Luminance: Energy perceived by observer
Brightness: Subjective descriptor
CIE-XYZ
RGB
CMY, CMYK
HSV, HSI, HLS
YIQ
YUV, YCbCr
LUT

6. RGB (Red, Green, Blue) Additive color space
Three primaries: red, green, and blue
Cannot always produce a color equivalent to any wavelength

7. HSI (Hue, Saturation, Intensity)
Similar to HSV (Hue, Saturation, Value)
Represents colors similarly how the human eye senses colors.

8. YUV (Luminance, Chrominance)
Similar to YIQ and YCbCr
Used for the PAL and SECAM broadcast television system
Amount of information needed to define a color is greatly reduced

9. CONVERSIONS
RGB TO YUV
Y =  .299R G B U = -.147R G B V =  .615R G B
RGB TO HSV
H = cos-1([(R-B)+(R-G)]/2*[(R-G)2+(R-B)(G-B)]1/2)
S = 1 – 3[min(R,G,B)]/(R+G+B)
V = (R+G+B)/3
Can we reduce the color space by using unsupervised dimension reduction techniques (like PCA)?
Can we use different domains?

10. AIM OF COLOR SEGMENTATION
For each object, find the most accurate subspaces of the color space to represent the object.
YUV seems the most promising color representation for our real time applicaiton.

11. FINDING THE SUBSPACES
First label images, then use supervised pattern recognition techniques. Most common ones:
C4.5
MLP
KNN

12. C4.5 Forms a decision tree for classification.
Uses the concept of information gain (effective decrease in entropy).

13. MLP (Multi-Layer Perceptron)
The MLP network is suited to a wide range of applications such as pattern classification and recognition, interpolation, prediction and forecasting.

14. KNN (K-Nearest Neighbor)
KNN is a simple algorithm that stores all available examples and classifies new instances of the example language based on a similarity measure.

15. CONDENSED KNN
Condensed KNN: We can reduce the training set by removing the samples that introduce no extra information to the system.

16. PCA (Principal Component Analysis)
PCA is a mathematical procedure that converts a number of possibly correlated variables into a hopefully smaller number of uncorrelated variables called principal components.

17. REGION BUILDING AND MERGING

18. RLE (Run Length Encoding)
RLE encodes multiple appearances of the same value.

19. REGION GROWING
This method depends on the satisfactory selection of a number of seed pixels.
This method may be performed before color segmentation.

20. REGION MERGING AND SPLITTING
Merging algorithms: in which neighboring regions are compared and merged if they are similar enough in some features.
Splitting Algorithms: in which large non-uniform regions are broken up into smaller regions which is hopefully uniform.

21. OBJECT RECOGNITION
OPPONENT GOAL
BEACON
BALL

22. CLASSIFICATION Already done by color segmentation.
Ball: The biggest blob with “ball color”,
Beacons: Two adjecentblobs with beacon colors, etc.
Unclassified blobs are discarded.
Each object is classified with a certainty.

23. TEMPLATE MATCHING Accomplished by using convolution or correlation.
Only works for translation of the template.
In case of rotation or size changes, it is ineffective.
Also fails for partial views of objects.

24. SANITY CHECK
A series of sanity check inspections are performed by the AIBO vision module to ensure the object classification is logically correct.
Ball cannot be above the goal,
Goals cannot be below the field,
There cannot be two balls, etc.

25. BOUNDING BOX CREATION
Requires affine transformations (translation, rotation, scaling)
Required for calculating distance and position information
The final products of the vision module are the bounding boxes of each visible object.

26. REFERENCES Cerberus RoboCup 2002 Team Report
rUNSWift RoboCup 2002 Team Report
NUBots RoboCup 2002 Team Report
CMPack RoboCup 2002 Team Report
MACHINE LEARNING, Mitchell
MACHINE VISION, Jain, Kasturi, Schunk

27. REFERENCES
Any Question?