Stixel-based Stereo Vision for Real-time ADAS

초록

Stereo camera has been researched steadily in field of Advanced Driver Assistance Systems(ADAS) and self-driving system to replace LIght Detection And Ranging(LIDAR) which is hard to commercialize. This paper suggests a system that uses stereo vision to detect obstacles in front of the vehicle in real-time. Particularly, Stixel lively used in Daimler is adopted. In addition, local matching with weighted least squares filter, limited v-disparity based free space and height constraint are used to decrease computational costs. Also in order to evaluate reliability of distance measurement using stereo camera, analysis was performed to compare the accuracy of the measured distance between mono-camera and stereo camera. Experiments were evaluated by using Daimler stereo dataset, similar to actual driving condition. Error of computed distance was evaluated by comparing the distance through Stixel with that of ground truth from ego-vehicle to object provided by Daimler. In case of mono vision, it was evaluated by using representative value based on the ground. As a result, an error measured by stereo camera was confirmed about 1% in the distance within 35 meters. In addition, suggested system satisfied real-time, 34fps in VGA image.