Visual Inertial Odometry with Pentafocal Geometric Constraints

Abstract

We present the sliding-window monocular visual inertial odometry that is accurate and robust to outliersby employing a new observation model grounded on the pentafocal geometric constraints. The previous approachesare dependent on the unknown 3D coordinates of the features to estimate the ego-motion. However, the inaccurate3D position of the features can lead to poor performance in motion estimation. To overcome these limitations,we utilize the pentafocal geometry relationship between five images as camera observation model, which makes itunnecessary to estimate the 3D position of the features. Furthermore, we apply the pentafocal constraints in the1-point random sample consensus (RANSAC) algorithm to find incorrect feature correspondences. We demonstratethe effectiveness of the proposed algorithm in two types of experiments: the KITTI driving scene dataset and theEuRoC micro aerial vehicle (MAV) flying dataset, both qualitatively and quantitatively. It shows more accurate stateestimation performance compared to the well-known stereo visual odometry algorithm and current state-of-the-artvisual inertial odometry methods.