Considering the current state in service-robotics, an expert is still necessary to add new tasks and execution behaviors by textual and error-prone programming. Under the consideration that humans typically execute same activities almost identical (or at least similar) and further combine simple behaviors to more complex activities, we follow the constitutive assumption that all complex behaviors are composed of a limited set of atomic behaviors. This work introduces a generic framework for spatialtemporal analysis and classification of arbitrary atomic behaviors. Therefore, we propose the combination of Self-Organizing Maps (SOM) and Probabilistic Graphical Models (PGM) in order to exploit the advantages of both concepts. In this work, we describe the essential methods of the framework briefly, whereas the datadriven training of the spatial-temporal model and the reasoning process are described in detail. In order to demonstrate the potential and to emphasize the high level of generalization and flexibility in real-world environments, the framework is evaluated in an exemplary scenario.