Color Segmentation

To find a colored dot in its search window, we first determine the pixel that has the smallest RGB-distance to the color of the dot's model. We assume that this pixel belongs to the dot and try to segment the remaining pixels from the background by analyzing a quadratic window with a side length of about twice the dot's diameter that is centered at the selected pixel. We use two methods for color segmentation. The first method that is illustrated in Figure 2 works in HSI color space [5]. Since we are looking for colorful dots, we apply a saturation mask and an intensity mask to the window. Only pixels that are saturated and neither too bright nor too dark are analyzed further. The final segmentation decision is done using the hue distance to the model's color. Working in HSI space has the advantage that the method is quite insensitive against changes in intensity. However, it can only be applied to saturated dots. If the model dot is not saturated or the HSI segmentation fails, we try to segment the dot in RGB color space. This backup method is needed, since some teams might use unsaturated markers, and also since saturated markers may appear as unsaturated dots when viewed in shadow or hot spots. Here, we use only the RGB-distance between the pixels color and the color of the model. All pixels of similar color are segmented as belonging to the dot.

**Figure 2:** Segmentation of colored dots in HSI space. Shown are (a) the original images, (b) the saturation/intensity masks, (c) and the segmented pixels, that have been selected using the hue distance to the model.
$\begin{figure} \begin{center} \begin{tabular}{ccc} \psfig{figure=seg1.eps,width... ...5ex} (c)\\ \\ not moving& & moving fast \end{tabular}\end{center}\end{figure}$

A quality measure is computed for the dot from the similarity to its model. If the size and the color are similar to the model, then the quality is high. The dot's position is updated if its quality is high enough and the dot is part of a valid robot, or if it is the ball. We estimate the position of the dot with sub-pixel resolution as the average location of the segmented pixels. If the quality is good, the size and average color of the dot are adapted slowly, as long as they do not deviate significantly from the initial model.

The segmentation does not explicitly take the shape of the dot into account. If enforces compact dots by searching many pixels in a small quadratic window.