To find a robot, all dots that belong to its model are searched for. If all dots can be localized, we compute the robot's quality from the degree of similarity to its model. We take into account the quality of the individual dots and their geometry, e.g. the distance between them.
For our robots, where three colinear equidistant colored markers are present, we check if the dots found fulfill this geometric constraint. If only one of the three dots is not found with sufficient quality, we can exploit the redundancy of our model. From the locations of the two dots found, we know where the third one should be and can check if it is really there.
We update the model of the robot only if the quality of the robot found is high enough. The position is adapted faster and more often than the geometry.
To prevent the tracking of non-robots, we count how frequently the quality of the robot found is low. If the robot is found for a certain time with low quality, then it is considered lost and a global search is performed to find it again.
If we lose objects, we assume that they are still at the position where we saw them last and signal this to the behavior control. If we lose the ball next to a robot that is near to the goal line, we assume that this robot occludes the ball and update its position together with the robot. This feature proved to be useful in penalty situations.