In-vehicle Pedestrian Detection Using Stereo Vision Technology

In the last 3 decades, between 4378 (in 2008) and 8090 (in 1979) pedestrians were killed each year in motor vehicle related crashes, representing 11% to 17% of the total roadway fatalities. Although the numbers of pedestrian deaths have been in decline steadily since 1980, their distributions have become more and more concentrated in urban areas. There is an urgent need to develop reliable pedestrian detection technologies that can warn drivers in time to take corrective actions to avoid collision with pedestrians. This paper presents the research findings of the development of an in-vehicle pedestrian detection system using stereo vision technology. Stereo vision images contain both color and depth (distance) information of each pixel, giving researchers the option to implement more efficient filtering algorithms to quickly reduce the regions of interest (ROI). The developed system consists of one pair of stereo vision cameras, one vision accelerator, and one Dual Quad-Core computer. A layered approach was implemented to systematically remove irrelevant pixel regions, reject non-pedestrian objects, and then using pattern matching techniques to identify and track pedestrian like objects. The developed system can recognize pedestrian like objects, and other objects such as ground, vehicles, buildings, trees, and tall structures. It was tested under day light and twilight conditions. The completed system can process videos at 7 to 10 Hz rates, detect pedestrian like objects up to 30 meters away while driving at speed of up to 35 mph, and achieved a 90% overall positive detection rate. The project was funded under the FHWA Exploratory Advanced research Program.