Perception and task-specific interpretation of dynamic environments represent key components of forthcoming intelligent systems. To master these abilities, methods are required that are capable of extracting relevant information from signals and combining them adequately to construct a semantically enriched model of the scene of interest. This contribution focusses on two aspects of this task. On the one hand, a method is presented to make an optimal selection from the available input data. On the other hand, an object-oriented environment model is proposed that enables a continual fusion of available knowledge with new sensor information and combines this with a memory model. All methods are based on Bayesian statistics in an objective degree-of-belief (DoB) interpretation. The application areas of these approaches are demonstrated with humanoid robots and autonomous vehicles.