Future intelligent vehicles will be equipped with cameras for driver information, warning and active assistance. Lane departure warning is at the market and more sophisticated systems for lane keeping on highways and distance control in slow traffic will follow. In order to further extend the capabilities of the vehicles' new eyes to everyday traffic, algorithms are being developed for understanding urban traffic scenes. In this contribution the approach is described to build an intelligent real-time vision system for this scenario. This includes stereo vision for depth-based obstacle detection and tracking, a framework for monocular detection and recognition of relevant objects including pedestrians and traffic signs and an attempt to realise such a system without the necessity of a super-computer in the trunk. The computational power in the demonstrator car UTA II (Urban Traffic Assistant) is currently limited to three 400 MHz dual-Pentium II PCs. For the covering abstract see ITRD E114174.
Abstract