Robust Initialization for Stereo Visual-Inertial SLAM with Polar-Based Gravity Estimation

Yifan Liu,Yuan Rao,Guanqi Qi,Hao Fan,Junyu Dong

DOI: https://doi.org/10.1109/icdis55630.2022.00067

2022-01-01

Abstract:Visual-Inertial initialization aims at estimating key parameters for Visual-Inertial Simultaneous Localization And Mapping (VI-SLAM) systems. However, most existing methods put more emphasis on making all estimations jointly consistent whereas ignoring the influence of gravity estimation on accelerometer bias estimation, resulting in unreliable parameter estimation. This hinders the performance of VI-SLAM under challenging environments. In this paper, we propose a novel stereo visual-inertial initialization method in a coarse-to-fine manner to estimate the system parameters accurately. Concretely, our method consists of four stages: 1) The stereo visual tracking with a track-then-verify strategy to generate robust matches for reliable vision-only state estimation; 2) The IMU preintegration with Euler integration for preprocessing raw IMU data and obtaining the inertial-only state estimation; 3) Loosely-coupled visual-inertial alignment for coarse-level parameter estimation with polar-based gravity estimation; 4) Tightly-coupled nonlinear visual-inertial optimization for fine-level parameter estimation. Comparative experiments on the EuRoC dataset show that our method achieves highly competitive performance against the state-of-the-art methods in gyroscope bias, accelerometer bias, and gravity direction estimation.