3D-model-based-vision for innercity driving scenes

The complexity of inner city traffic areas presents a considerable challenge for driver support based on machine vision. This is due to a generally high density of different objects in the scene, sharing varying spatio-temporal relations, as well as difficult imaging conditions that complicate image evaluation. Road borders, vehicles, traffic signs, and other infrastructural objects have to be detected and localized within the scene in order to robustly and accurately assess traffic situations around the vehicle. In this context, model-based vision allows to select relevant image structures exploiting a-priori knowledge given by geometric object models for the scene structure and properties of objects. We report a model-based approach which uses 3D object models and a-priori knowledge about typical positions of traffic signs and vehicles w.r.t. the road in order to detect and track such objects within image sequences recorded from within a driving vehicle. The method is integrated into a machine-vision-based system used to track lane boundaries and lamp posts located next to road borders. The multitude of different objects related to each other facilitate consistency checks and thus increase the robustness of the overall system.