As a computer science research intern, Yi spent the summer of 2005 at Bowdoin College working with Professor Eric Chown on studying and improving the vision system of the Aibos--robot dogs we program to play soccer games against other universities and colleges.
To perform color blob, line, or edge identification, a picture captured by the Aibo's camera need to be thresholded first--every pixel in the raw image has one of 16 million values, that is too much information so those values need to be turned into basic colors like "blue" or "red." The vision system the Aibo's used previously used simple ranges in the color space to define a color (think of placing a big rectangle over the space of possible values) and thresholded pictures according to the calibrated those ranges. The old threshold algorithm was fast but it is not accurate enough, because its definitions of colors are just cubes in the color space where the real "shapes" of the colors could be virtually anything. Yi rewrote the threshold algorithm so that it is more accurate since it uses color tables to map the raw values to a color. The new threshold algorithm also has some useful new features such as allowing a pixel to be labeled with multiple colors. Yi also created a set of tools (including tools for sample collecting, analyzing and generating color tables) that facilitates the process of creating color tables use by the new algorithm.
This project is continuing as an independent study during the school year. Yi has improved the speed of the threshold process by making use of the Color Detection hardware on the Aibos. The new vision system loads the color tables created by Yi into memory and performs the thresholding with algorithms that are built in to the Aibo's hardware. We are hoping that the performance of the vision system will further improve and help the Aibo soccer team improve its performance.