Fast Object Tracking

このブログ：Fast Object Detecting のアルゴリズム概要が公開された！！！！にFast Object TrackingのBrad Hayesさんが物体認識のアルゴリズムの概要を公開されました。のような内容が書かれて、無理やりここにコピーしといて、時間が有れば、確認したいと思います。

Tennis Ball Tracking

1. Load Texture Patch and detect average color value
2. Remove pixels (set to 0) that are more than THRESHOLD distance away from detected color average of target or below some NOISE_THRESHOLD
3. Compute Integral Image on the result
4. Perform constricting window algorithm to detect densest square area of energy in the integral image.
5. Return coordinates (x,y,l) of square: Top left corner coords (x,y) and square side length l.
   

Code Sample

Touch/Impact Tracking

1. Iterate through frame
2. If the current pixel intensity is below the specified threshold: Set all channels at the current pixel that are below NOISE_THRESHOLD to 0 and move on to the next pixel
3. Call recursive "expand" function on current pixel location that searches radially outwards from the current point for connected pixels of sufficient intensity. These pixel locations and values are added to linked lists and then are set to 0 in the image to prevent infinite loops.
4. Calculate the centroid (x,y) of a given collection of connected points as well as the distance from the center to the furthest pixel in the blob (define as 'r')
5. Result: Circular blob at coordinates x,y with radius r.