Convergent Active Stereo: Investigating Depth Perception with a Robotic Binocular Camera System

Abstract

While extensive research exists on stereo disparity estimation algorithms for fixed parallel systems, there is limited exploration of active convergent stereo cameras. This thesis investigates the differences in depth perception between a robot head mimicking human eye movement with the ability to fixate and a static parallel setup. A novel convergent stereo algorithm, which is capable of searching for correspondences with diagonal epipolar lines, is proposed. To evaluate the algorithm, the first natural image convergent stereo dataset was created, which includes ground truth for parallel views, parallel images of each scene, and images of fixations across the scene. An in-depth analysis reveals the advantages and limitations of both systems, highlighting the geometric benefits of active convergent stereo, especially in challenging scenes. The results from these findings can contribute to further the development of more effective vision-based depth perception system designs for humanoid robots.