Dynamic Initialization Problem

[QingfengLi-hit]

The current initialization scheme requires static conditions. If the static conditions are not good, it will drift away later. I plan to use SFM to solve , but with this method, the covariance in the sliding window cannot be obtained, and only the result of the state(ba, bg, pose,v) can be used.
Is there any good initialization scheme for msckf?

[goldbattle]

Here are some papers that might get you started, hope this helps.

Closed-form solution of visual-inertial structure from motion
Agostino Martinelli
https://hal.archives-ouvertes.fr/hal-00905881/document

Visual-Inertial Monocular SLAM with Map Reuse
Raul Mur-Artal, Juan D. Tardos
https://arxiv.org/abs/1610.05949

Robust initialization of monocular visual-inertial estimation on aerial robots
Tong Qin; Shaojie Shen
https://ieeexplore.ieee.org/document/8206284

Revisiting visual-inertial structure from motion for odometry and SLAM initialization
Georgios Evangelidis, Branislav Micusik
https://arxiv.org/abs/2006.06017

An Analytical Solution to the IMU Initialization Problem for Visual-Inertial Systems
David Zuñiga-Noël, Francisco-Angel Moreno, Javier Gonzalez-Jimenez
https://arxiv.org/abs/2103.03389