Interpretation of Spatio-temporal Relations in Real-Time and Dynamic Environments Andrea Miene and Ubbo Visser TZI - Center for Computing Technologies University of Bremen Universit¨ atsallee 21-23 D-28359 Bremen, Germany {andrea|visser}@tzi.de Abstract. With the more sophisticated abilities of teams within the simulation league high level online functions become more and more attractive. Last year we proposed an approach to recognize the opponents strategy and developed the online coach accordingly. However, this approach gives only information about the entire team and is not able to detect significant situations (e.g. double pass, standard situations). In this paper we describe a new method which describes spatio-temporal relations between objects. This approach is able to track the ob- jects and therefore the relations between them online so that we are able to in- terpret situations over time during the game. This enables us to detect the above mentioned situations. We can implement this in the online coach in order to en- rich our team with high level functions. This new method is domain independent. 1 Motivation The online coach is still the most effective instrument to analyze the opponent, because it can obtain all information about the simulated environment. Therefore, it is important to continue the development of the online coach which we used for the Virtual Werder team in the RoboCup 2000 tournament. In [Visser et al., 2001] we describe how the coach determines the opponent tactical formation with a neural network and how it is able to change a team formation during a match. We showed that it makes sense to recognize strategies and change the own team accordingly. However, our approach relies on information about the opponent’s team in total and is therefore not able to recognize and/or predict ’local’ situations. We believe that the detection of the opponents behavior in smaller areas, e.g. a double pass or a standard situation would help to find the appropriate countermeasures. The online coach is the optimal player for the collection of this kind of information and it is obvious that the coach should be able to process the data and find the appropriate tactic for the own team. Also, the analysis should be available online as the developed methods should be able to function in a real-time environment. In this paper we describe a new method that is able to track moving objects in real time. The idea is to detect spatio-temporal relations between objects (players and ball) in a first step and then learn from this observations whether there is a repeating pattern, e.g. an attack over the wings with a pass onto the penalty point. A. Birk, S. Coradeschi, and S. Tadokoro (Eds.): RoboCup 2001, LNAI 2377, pp. 441–446, 2002. c  Springer-Verlag Berlin Heidelberg 2002