Knowledge-Driven World Modeling

In order to allow safe operation and good performance, robots need an accurate model of their environment. Such a world model is constructed from sensor detections and typically contains information about positions and velocities of surrounding objects. This paper proposes how prior knowledge about those objects can be used to improve the performance of a world model, which is implemented by a Multiple Hypothesis Filter (MHF). More specifically, knowledge about object dynamics, expected locations, relations between object classes and detector characteristics is incorporated in the probabilistic models of the MHF. The results of simulations confirmed the potential of incorporating such object knowledge in a world model.