Stereo Visual Inertial Odometry for Unmanned Aerial Vehicle Autonomous Flight

Abstract

Visual–inertial SLAM system is very popular in the near decade for the navigation of unmanned aerial vehicle (UAV) system, because it is effective in the environments without the Global Position System (GPS). Due to size and weight constraints, only inexpensive and small sensors can be used. Therefore, there are still challenges in the computational efficiency and robustness of MAV autonomous flight algorithm. We present S-VIO: an optimization-based stereo visual-inertial odometry. Our approach starts with inertial measurement units (IMU) pre-integration, in which IMU measurements is accumulated between several frames using measurement pre-integration. After the initial state estimation converges, a highly precision stereo vision-inertial odometry is obtained by fusing IMU measurements and feature observations. Our approach is validated on the EuRoC MAV datasets. Experimental results prove that our S-VIO has higher accuracy and robustness than the most advanced visual-inertial fusion methods in some challenging situations.