フィリップ フェベ, 井上誠喜
Handling of Human Motion
Abstract:Coming in ATR, I knew that my studies here would be around motion tracking. I did not know a lot about it, and I had tried to get some information about it before coming to Japan. But I just had a look at some results in some books and that was all, because this domain is a very large one.
During the first month here, I learned to use my computer, to program in C some simple image processing routines and read about algorithms that could be of use in order to perform motion tracking. I understood after a short time in ATR that such motion tracking was part of a more larger area which was computer vision. Attending various talks about the subject, I could see how important such a domain was.
As my study was about handling of human motion, I decided first to try some methods based on color detection. Because the interesting parts to be tracked on human body are especially the face and the hands. It seemed that recognizing skin color would not be too heavy at a computational level. Finally, I was able to get some zones in the image, these zones being either the face, either the hands or some noise.
At this moment, I had to make some choices for the recognition part of the processing. And it showed up that this part was the crucial one. Depending highly on the conditions of use. I had no idea if I could make it real time or not.
That's why I went to video programming. I needed to get used to video in order to make some tests on it. And that's what gave birth to the template matching algorithm. I tried this algorithm because it was quite a simple one which was based on difference between frame and would bring me to video. As this method was performing better than expected, I spent time testing it and trying to improve it.
Then I came back to color and video. And I applied the color detection on video and observed the frame rate. It seemed clear that the processing was not so quick and that therefore, there would be a problem for real time, because the recognition requires further treatment. And also how to perform recognition remained unclear to me. It appeared that the recognition was linked with the conditions of use of this color detection, because reliable recognition would mean lots of calculations.
It is at this moment that I began to go in the studio. Because the tracking algorithms were performed to run there. So they had to run there. Consequently, I had to make some changes on the programs, because the machines were differents. It is during that time that I had the idea of making a boundary extraction of a person in the studio so that I could get high level information about the shape. As it will be discussed more in the following pages, I gave up this idea after a while.
From that time, I began working on possible applications for template matching: one in the studio, and the other being a motion database making.
Consequently, I will speak about the three methods I had the idea to apply during my stay at ATR: the color detection, the boundary extraction and the template matching. I finish with template matching because the further applications use this method.