A Detailed Description of Direct Stereo Visual Odometry Based on Lines

Abstract

In this paper, we propose a direct stereo visual odometry method which uses vertical lines to estimate consecutive camera poses. Therefore, it is well suited for poorly textured indoor environments where point-based methods may fail. We introduce a fast line segment detector and matcher detecting vertical lines, which occur frequently in man-made environments.We estimate the pose of the camera by directly minimizing the photometric error of the patches around the detected lines. In cases where not sufficient lines could be detected, point features are used as fallback solution. As our algorithm runs in real-time, it is well suited for robotics and augmented reality applications. In our experiments, we show that our algorithm outperforms state-of-the-art methods on poorly textured indoor scenes and delivers comparable results on well textured outdoor scenes.