Although many different vision algorithms and systems have been developed so far, integration into a complex intelligent control architecture of a mobile robot is in most cases an open problem. In this paper we describe the integration of different vision based behaviours into our architecture for sensorimotor systems. By means of different scenarios like person tracking and searching of different objects, the structure of the vision system and the interaction with the overall architecture is explained. Especially the interaction of vision based modules with the task level control and the symbolic world model is an important topic. The architecture is successfully used on different mobile robots in natural indoor environments.