Abstract
Automobile technologies have been challenging to realization of safe driving by the employment of a computer vision system. Various techniques have been developed for preventing car accidents based on so-called car vision technologies by analyzing road images captured by the cameras mounted on a car. Traffic accidents often occur at junctions compared to motor ways or ordinary roads. We therefore propose in this paper a technique for detecting pedestrians on a zebra crossing near a junction from the video images provided by a camera on a car. The detection is performed by recovering backgrounds from the video images and separating foregrounds from the backgrounds. Performance of the proposed technique was examined by the experiment employing car video images at some urban junctions. The technique detected 82% in average of a pedestrian's shape correctly when 1% of the background was incorrectly extracted as part of the pedestrian.
