Dynamic event description for a mobile robot is a very challenging task. Event description is based on an ontology. Recent research work tries to build up this ontology autonomously. Therefore this work proposes a first step towards that: a system for dynamic learning of scene descriptions. By enabling the robot to recognize the context where it is at the moment, it is possible to significantly increase the performance of object recognition systems and the creation of an ontology can be simplified. Based on the concept of the self-organizing map a system that can continuously recognize and assimilate scene descriptions into its knowledge base was developed. Clustering techniques are used to categorize the scenes and provide an interface to an context supported object recognition system and further to an ontology.