Accuracy Analysis of Photogrammetric Stereo Visual Odometry according to Imaging Geometry

초록

Visual Odometry (VO) is a technique to estimate platform motion using an image sequence. This is one of promising image processing techniques because it uses only images at relatively low cost. However, for the same reason, this technique’s accuracy becomes sensitive to imaging geometry. In this study, we analyze the accuracy with two conditions: Field of View (FOV) of images used for experiments, and baseline, the distance between two imaging locations. The Photogrammetric Stereo Visual Odometry (PSVO) developed in this paper performs feature extraction and matching, feature optimization, and photogrammetric motion estimation. For experiments, we used a dataset provided by Karlsruhe Institute of Technology and Toyota technological Institute (KITTI) community and a dataset acquired with our platform. Our dataset is called Inha dataset and has a smaller FOV and a longer moving distance per frame than KITTI dataset. We compared the results depending on imaging geometry and performed visual inspection of feature matching and accuracy verification. In the experimental results, as FOV decreases, the estimation errors tended to increase. Also, the longer moving distance per frame, the worse performance of outlier filtering, especially Random Sample Consensus (RANSAC). Through these experiments, we observed that not only the status of features, but also imaging geometry was critical factor and the general RANSAC filtering was not suitable for VO. This paper proposes the importance of imaging geometry and feature optimization in VO.