Current driver assist systems feature basic lane recognition abilities when applied under controlled conditions, e.g. driving on a motorway at daylight. Urban intersection scenarios, however, are still hard to understand. The reason for this is multifaceted: Innercity crossroads connect highly frequented roads, thus traffic participants block the view of onboard sensors. Furthermore the complex lane geometry accounts for a high-dimensional parameter space.

This work will evaluate and extend existing approaches for crossroad detection and recognition. The goal is to understand complicated intersection geometries, based solely on observations from on-board video cameras. To achieve robust estimation, multiple different visual cues have to be integrated into the recognition process. A system capable of estimating the ego-lane and crossing lanes at intersections could then be used for driver assistance in complicated situations or even autonomous negotiating.