Most approaches to detection and classification of human activity deal with observing individual persons. However, people often tend to organize into groups to achieve certain goals, and human activity is sometimes more readily defined and observed in the context of whole group, where the activity is coordinated among its members. An excellent example of this are team sports, which can provide valuable test ground for development of methods for analysis of coordinated group activity. We used basketball play in this work and developed a probabilistic model of a team play, which is based on the detection of key events in the team behavior. The model is based on expert coach knowledge and has been used to assess the team performance in three different types of basketball offense, based on trajectories of all players, obtained by wholebody tracker. Results show that our high-level behaviour model may be used both for activity recognition and performance evaluation in certain basketball...