A metric for pose Quality and consistency

[aubino]

Hello everyone.
I'm using Openvins in a robotic application and my environnement is verry constrained and
sometimes lacks visual texture.
In those conditions the estimator can often diverge and rightfully so because the observation is wrong.
I'm trying to come up with a metric to give me how much i can trust the system's estimation,
because divergences in audometry can cause my platform to behave eratically.
I have tried indexing the trust on the estimation's variance (sqrt(cov)) but most of the time it does not grow fast enough.
Does anyone have an idea ?