Safety and Precision of Spatial Context Models for Autonomous Systems

A safe (i.e. collision-free) and efficient movement of a mobile autonomous system requires a correct assessment of the vehiclesʼ environment. In many systems, the spatial context is represented by means of an occupancy grid, that is, a partitioning of a two-dimensional area into finitely many cells where each cell maintains a probabilistic estimation of being occupied by some object. Often, this representation is complemented by dedicated information regarding the detected objects. We instantiate a generic framework based on a formal notion of abstraction for characterising the safety and precision of such kind of spatial context models. In this approach, the sensory perception corresponds to an abstraction function, and the concretisation function reflects the loss of information with respect to the real environment. We show that this method is able to derive both qualitative and quantitative measures for different variants of spatial context models.