At any given time, most objects move slowly on the field and only some objects might move fast. This allows to predict the positions of dots from the difference of the last two positions. We use linear prediction clipped at the borders of the field (see Figure 1). If an object moves fast, this prediction is not very accurate. Therefore, we need a larger search window for this situation. However, if the object stands still, the prediction will be very good and a small search window suffices.
We developed an algorithm for dynamic adjustment of the search window sizes.
If we find the dot within its window, we slowly decrease the window's side length for the next frame.
If the search is not successful, we increase the size by a factor of two and search again, as illustrated for the ball in
Figure 1.
The size is always limited between a minimal (e.g. 16
16) and a maximal value (e.g. 128128) and the searched region is clipped by the rectangle containing the field. During regular play, the average size of the search windows is close to the minimal value. Large windows are only necessary, when objects move very fast (e.g. the ball has been kicked) or dots cannot be segmented (e.g. due to the lighting or occlusions).
One problem when enlarging the search windows for the robot's dots is that one can find dots of the same color that belong to different robots. To prevent this, we use two heuristics. First, we remove dots of found robots from the image by ``painting" them black. Second, we enlarge a search window only after all dots have been searched for in their small window.