A Space Variant Mapping Architecture for Reliable Car Segmentation

Abstract

Side-impact crashes have now become more important than head-on crashes, probably reflecting improvements in protecting occupants. Overtaking scenarios are one of the most dangerous situations in driving. This paper is concerned with a vision-based system on the rear-view mirror for safety in overtaking scenarios. A bio-inspired algorithm segments overtaking vehicles using motion information and rigid-body-motion criterion. The overtaking scene in the rear-view mirror is distorted due to perspective. Therefore we need to devise a way of reducing the distortion effect in order to enhance the segmentation capabilities of a vision system based on motion detection. We adopt a space variant mapping strategy. In this paper we describe a computing architecture that finely pipelines all the processing stages to achieve reliable high frame-rate processing.