On-road object validation using surface normal vectors in stereo vision for ADAS; [표면 벡터를 를 이용한 스테레 오 비전 에서의 물체 검출 검증]

초록

The purpose of this paper is to improve the accuracy of object detection by validating the objects detected using surface vectors in stereo vision for object detection in autonomous vehicles or robots. Conventional stereo vision algorithms have a problem of increasing disparity errors between two images by simplifying correspondence matching processes for the purpose of real-time processing. This paper proposes a method to correct the detection result using the surface vector to reduce the false positives due to erroneous disparity images. The proposed method uses Stixel to reduce computation cost when searching for objects in the disparity image, calculates the surface vectors to minimize the false positives of objects due to degradation in disparity image, evaluates the disparity confidence, and then verifies the object. The surface vector is locally computed from the stixel-based ROI to maintain realtime processing. Experiments were carried out using real environment images to verify the performance of the proposed method. Its accuracy was improved by more than 7% compared to the existing stixel methods and the processing time of the surface vector module was less than 6 ms in a PC platform. © ICROS 2018.

키워드