Open mazatov opened 4 years ago
As I understand larget std
should result in larger covariance
and hence smaller gating_distance
but they don't seem to have much of an impact.
_std_weight_position
and _std_weight_velocity
are scalars multiplied with the height of the bounding box, a solid way to initialize the covariance matrix given that images have no depth information.
Making these values larger will increase uncertainty and therefore decrease the gating distance.
However, be aware that during the update step:
new_covariance = covariance - np.linalg.multi_dot((
kalman_gain, projected_cov, kalman_gain.T))
it is likely that the covariance matrix is significantly reduced due to the small innovation. This will not change the gating distance since the gating distance is computed before the update step, but it is worth noting. However, you have the right intuition. What you are doing should work.
I want to tune
_std_weight_position
and_std_weight_velocity
for my problem. Wonder if someone can provide some intuition for adjusting these parameters.I tried to change the values from the original
1/20
to 1/2 and to 10 and really haven't noticed difference in the resutls.Thanks!