Identification
Team G
Project 15
Team members
Lajos Rodek Marcin Rogucki Mircea Nanu Selman Kulac Zsolt Husz
-
Szeged, Hungary
Lodz, Poland
Timisoara, Romania
Ankara, Turkey
Timisoara, Romania
Lajos Rodek
Sign recognition ideas
Sign library preparation
Presentation
Lots of laughing
Marcin Rogucki
Sign recognition coding
Sign recognition ideas
Sign detection ideas
Presentation
Mircea Nanu
Sign detection ideas
Sign detection coding
Web page preparation
Moral support and
jokes
Selman Kulac
Gathering sign images
General ideas
Presentation
Zsolt Husz
Sign detection coding
Sign detection ideas
Picture acquisition
Many, many testing
Our goal
Assumptions
No human interaction
No preprocessing of the image
Flexible handling of images
Image is not rotated by more than 30
degrees
Images can contain any number of signs or no
signs at all
Only daylight images are taken
At most of a sign may be covered
No background constrains / limitations
Sign detection 1
Signs features:
Well defined colors with high saturation
They are rather homogenous
Sharp contours
Known basic shapes
Allowed colors:
Red, blue (dominant colors)
Yellow
Green (very rare)
White, black (found mostly inside of signs)
Sign detection 2
Main steps:
Edge detection (3 by 3 Sobel)
Converting image to HSV color space
Reducing number of colors
Segmentation relying on the color
Marking probable signs with boundary
boxes
Joining adjacent regions
Removing background
Sign detection 3
Input
Conversion
to grayscale
Sobel
Conversion
to HSV
Color
detection
Border
extraction
Region
joining
Output
Region
extension
Region
database
Sign recognition 1
Input:
Picture containing at most one sign
(subrange of the original image) with
eliminated background
Sign templates and names
Output:
Sign name in case it is a traffic sign
Localization on the image
Sign recognition 2
Tasks:
Detecting the shape of a sign
Finding corners if necessary
Transforming the shape
(Perspective/rotation Facing/upright)
Color unification
Comparison with templates
Sign recognition 3
Detecting the shape:
Building a chain code
Computing angles between vectors
Checking number of the corners
Defining a shape
(triangle,square,circle)
Sign recognition 4
Finding corners:
Charged particles
based approach
Particles run away from
each other and locate
corners as furthest
possible points in the
figure
Sign recognition 5
Transforming the sign:
Inverse texture mapping according to the
corners and shape
Sign recognition 6
Color unification:
Simplifying colors depending on similarity
Allowed colors:
Red, green, blue, yellow, white, black, background (pink)
Computing a histogram
Sign recognition 7
Comparison with a template:
Normalized histograms are compared
resulting in a RMS measure
Raster pictures are compared pixel by
pixel
Probability based decision
Results 1
Results 2
Results 3
Achievements
Everything works fine
Every team member is happy
Signs are detected and recognized
correctly in most cases
All assumptions are met
Works even in unusual cases (e.g. night
pictures)
Future improvements
Better reliability with fast motion blurring
More independency with illumination
Robustness on sign detection (fine-tuning
the heuristically adopted constrains)
Better library templates
Speed-ups
Adaptation for a sequence of images
References
Intel, Intel Image Processing Library, Reference
Manual, 2000, http://developer.intel.com
Intel, Open Computer Vision Library, Reference
Manual, 2001, http://developer.intel.com
D. A. Forsyth, J. Ponce, Computer Vision: A
Modern Approach, Prentice Hall, 2003
George Stockman, Linda G. Shapiro, Computer
Vision, Prentice Hall, 2001