Monocular visual inertial system based on semi-direct visual odometry

Aiming at the problem of low attitude estimation accuracy of monocular visual system, a monocular visual inertial system based on semi-direct visual odometry (SVO) is proposed. The system uses an error-state Kalman filter to fuse visual information and inertial navigation information in a loosely coupled manner. SVO is used to process image data in the visual front-end, and the SVO keyframe selection strategy is modified to solve the problem of attitude estimation failure due to relocation failure in the front-view scene of the camera. The proposed algorithm is validated by using the MH 01 subset of the EuRoc dataset. The experiment shows that the monocular visual inertial system based on the SVO can effectively improve the accuracy of attitude estimation. Compared with the monocular SVO algorithm, the position error of the proposed algorithm is reduced by 13.4%, the error of the rotation angle is reduced by 30%.