Abstract: Aiming at the problems of low localization accuracy as well as poor robustness of underwater visual navigation in weak texture environments, this paper proposes an underwater monocular visual inertial odometry based on the sparse direct method. The method is based on the assumption of pixel gray scale invariance, and estimates the camera position by optimizing the photometric error, avoids the complex process of feature point extraction and matching, thus improves the real-time and robustness of navigation, while combines the data from the inertial measurement unit (IMU) and uses error state Kalman filter (ESKF) for data fusion to further reduce the error, in order to improve the stability of navigation of autonomous underwater vehicle (AUV) in underwater complex environments. The stability and accuracy of navigation in underwater complex environments are improved. The experimental results show that the error reaches the centimeter level and is reduced compared with the vision-only algorithm, which proves that the system can effectively fuse vision and inertial information, and has high accuracy and robustness in the field of underwater navigation.

Key words: sparse direct method, autonomous underwater vehicle, inertial measurement unit, visual inertial odometry, error state Kalman filter