About Observation Centric update paradigm

[rslim97]

Hi Noah, I'm wondering if the paradigm of observation centric re-update with online-smoothing can be applied to other domains, such as KF tracking in the BEV plane? For the context, I'm using lidar detections to detect objects but the motion model sucks, so I hope with observation centric update it can smoothen things out?

[noahcao]

From a high-level perspective, my guess is - Yes. As whatever the representation is the concept behind observation centric re-update with online-smoothing is the same. I am not sure what stucks means here, can you look into the KF parameters to ensure that they evolve as expected?

[rslim97]

Hi Noah, what I mean is the motion model I'm using is a constant velocity one to model nonlinear motion of persons. Pardon me to bring this question up: what is the difference between calling an KF update with something like kf.update(None,None) and just doing nothing, i.e. doing no updates to the KF? also do you mean to eye on the evolution of KF parameters to see it makes sense?

[noahcao]

You can refer to the implementation of kf.update here:

OC_SORT/trackers/ocsort_tracker/kalmanfilter.py

Line 465 in 7a390a5

if z is None:

Simply speaking, by calling kf.update(None,None), you will refresh the observed status of the KF and also the related variables maintained.

You can check the KF parameters maintained to see if the variance (vectorized) of the velocity is smoothened or not.