Stixels estimation through stereo matching of road scenes

Abstract

Recently, Stixel-world, a medium level representation of road scene components has been introduced. The existing stix-els estimation approaches are separated from a depth estimation process, or they directly make use of stereo images and only compute stixels without producing per-pixel depth information. For road scenes, however, many machine vision tasks require both per-pixel depth information and the higher-level representation of it. This paper presents a combined process of stixels estimation and stereo matching process. The proposed method generates per-pixel depth information and stixels for both the ground surface and obstacles, at the same time. We have modified a multi-path line-optimization process of the stereo matching algorithm to produce multiple stixels of the ground and obstacle segments for each image column. Experimental results show that the proposed algorithm estimates stixels more accurately than the existing algorithm, and it also produces high-quality dense depth information, at the same time. © 2014 ACM.