Rapid Initialization using Relative Pose Constraints in Stereo Visual-Inertial Odometry

Abstract

In this paper, we propose an initialization method to bootstrap a visual-inertial odometry (VIO) without any prior information using relative constraints computed by a stereo camera. Many works on VIO start their algorithm from a static state or a rough initial guess since an accurate initial velocity and attitude are not available at the early stage. Our approach estimates a relative pose of consecutive camera frames from a stereo visual odometry (VO), then formulates least-square problems from the IMU preintegration and the pose constraint. This recovers the initial velocity, attitude with respect to the gravity, and biases of an IMU. Also, we build a framework in which a tightly-coupled filtering-based VIO is booted by the proposed initialization method. We show that the initial state can be estimated with a 6-DOF motion in the simulated environment. Also, our experimental results reveal that 5 seconds with 20 fps stereo images are enough to initialize the filtering-based VIO in the public MAV dataset.