Fast Monocular Visual-Inertial Initialization with an Improved Iterative Strategy

The initialization process has a great effect on the performance of the monocular visual inertial simultaneous localization and mapping (VI-SLAM) system. The initial estimation is usually solved by least squares such as the Gauss-Newton (G-N) algorithm, but the large iteration increment might lead to the slow convergence or even divergence. In order to solve this problem, an improved iterative strategy for initial estimation is proposed. The methodology of our initialization can be divided into four steps: Firstly, the pure visual ORB-SLAM model is utilized to make all variables observable. Secondly, the IMU preintegration technology is adopted for IMU-camera frequency alignment at the same time with key frame generation. Thirdly, an improved iterative strategy which is based on the trust region is introduced for the gyroscope bias estimation as well as the gravity direction is refined. Finally, the accelerometer bias and visual scale are estimated on the basis of previous estimations. Experimental results on the public datasets show that the estimation of initial values can be converged faster, as well as the velocity and pose of sensor suite can be estimated more accurately than the original method.