Right Invariant SE2 (3) - EKF for Relative Navigation in Learning-based Visual Inertial Odometry

The visual inertial odometry(VIO) has shown great potential localization ability in robots' autonomy tasks. Although the learning-based VIO method shows promising results as it is robust to different lighting conditions without sensors calibration. It is not necessary to learn the kinematics of IMU as its model is well studied. In this work, we propose a learning based VIO framework which uses a right invariant SE2 (3)-EKF to combine the measurements from IMU and relative poses from learning-based visual odometry. The novel SE2 (3) - EKF defines the right invariant error of the relative navigation model on matrix Lie group and derives the associated error state differential equations. As the relative pose of the learning-based VO is compatible with the right invariant error, the consistence issue of tradition EKF is overcame. The experiments on the KITTI datasets have shown the superior performance of the proposed method compared to the other learning-based methods.