On the Use of Motion Concepts for Top-Down Control in Traffic Scenes

Abstract

We showed how to exploit motion concepts associated with verbs of locomotion for top-down control in traffic scenes. Two kinds of constraints could be derived: spatial constraints through knowledge about the applicability of motion concepts, and motion constraints through knowledge about typical motion. We proposed to compute motion constraints using a spatio-temporal buffer as a shared representation for bottom-up and top-down processes. Within the buffer motion concepts are expressed as typicality distributions from which predictions about object motion can be derived. A local prediction algorithm allows for the computation of search areas for low-level motion analysis. A low-level motion representation based on spatio-temporal Gabor cells is well suited for the integration of this kind of top-down information. We presented an example where this procedure has been implemented. Using top-down guidance, the complexity of computation could be reduced significantly. Instead of analyzing the whole scene at the same level of detail, 1) only a small area could be chosen for an analysis and 2) the analysis could be focussed on specific spatio-temporal behavior within the area of interest.