Why dont inflate the covariance of position when recovering the initialization covariance?

[TianzeXie]

Why dont inflate the covariance of position but the other state like orientation velocity and other two bias?

[goldbattle]

Hi, could you expand on what exactly you are referring to? Thanks!

…

On Tue, Mar 19, 2024 at 8:35 PM Tianze ***@***.***> wrote: Why dont inflate the covariance of position but the other state like orientation velocity and other two bias? — Reply to this email directly, view it on GitHub <#433>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAQ6TYXLIA5Q5WQATAS6RDTYZD7ZRAVCNFSM6AAAAABE6XAWY2VHI2DSMVQWIX3LMV43ASLTON2WKOZSGE4TMNRQGIZDMOI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[TianzeXie]

here:

open_vins/ov_init/src/dynamic/DynamicInitializer.cpp

Line 1076 in 17b73cf

covariance.block(0, 0, 3, 3) *= params.init_dyn_inflation_orientation;

In my opinion, the reason is the covariance of position donot impact the other states because we dont use position to calculate any other state.

[goldbattle]

Hi, the global position and yaw are unobservable for visual-inertial systems, so there is no need to inflate these parameters are we treat them as 100% known. You can see this in the prior we assign (can also see we put a prior on bias as it is near unobservable for small accel and gyo changes):

open_vins/ov_init/src/dynamic/DynamicInitializer.cpp

Lines 696 to 699 in 17b73cf

	Eigen::MatrixXd prior_Info = Eigen::MatrixXd::Identity(10, 10);
	prior_Info.block(0, 0, 4, 4) *= 1.0 / std::pow(1e-5, 2); // 4dof unobservable yaw and position
	prior_Info.block(4, 4, 3, 3) *= 1.0 / std::pow(0.05, 2); // bias_g prior
	prior_Info.block(7, 7, 3, 3) *= 1.0 / std::pow(0.10, 2); // bias_a prior

We just inflate the initial orientation as its roll and pitch are observable. I found in simulation that the covariance was overconfident when using short initialization windows. This is attributed to the fact that we need full 3d motion (orientation and non-constant velocity) to fully recover our states. Hope this helps.