Currently, the ball and the players are recognized on the basis of a
certain color space and manually adjusted colour bounds for the
objects of interest. However, the colour bounds heavily depend on the
lighting conditions. This project aims at enhancing the vision
component of the KiRo software such that the systems auto-adapts to
changing lighting conditions.
Part A -
Automatic Calibration of Color Classifcator
Approach:
Find the color bounds automatically. Don't use any starting hypotheses. Vary
systematically the color bounds for the the relevant colors in the
table soccer game. Find an evaluation function which judges whether
the color bounds are adjusted optimally for a certain
color. Exploit table soccer domain knowledge, e.g. number of men per
rod, position of rods, only one ball is in play...
Evaluation:
Create at least 4 short video log files (each 15 seconds) under different lighting
conditions. (Room lights switched on/off, different settings for
camera gain/shutter speed). Make screenshots of vision window - with
the result of automatic calibration and the result of manual
calibration. Comment on the screenshots. For each color measure the
time (average and maximum time) needed by the method.
Delivery:
Line Calibration: 04.12.03
Rod Calibration: 18.12.03
Ball Calibration: 08.01.04
Part B -
Dynamic Adjustment of Color Classificator
Approach:
Adapt dynamically the color classificator (found in part A) when lighting conditions
change. Use the evaluation function found in part A. Find a heuristic
to speed up the search back to the optimum.
Evaluation:
Create one long video log file (one minute) while altering the lighting conditions.
Stop the video at predefined points in time and make screenshots -
with the dynamic adjustment switched on and switched off. Comment on
the screenshots. Measure the time (average and maximum time) needed
for finding back to the optimum.
Delivery:
12.02.04
Part C -
Object Recognition without explicit color classificator?
Approach:
Implement an auto adjusting vision system following the idea proposed
in: A Real-Time Auto-Adjusting Vision System for Robotic
Soccer by M. Jüngel, J. Hoffmann and M. Lötzsch
Evaluation:
Compare the approach to the method developed in Parts A+B. Document
differences in quality and processing time.
Institut für
Informatik, Universität
Freiburg