Simple Gaussian Mixture Filter (Tracker) argoverse_simple_gmf_tracker

Summary

This is a modified version of the Open Argoverse CBGS-KF Tracker used for the second place submission for the 2020 Argoverse 3D Tracking Competition. For a more detailed explanation, please refer to the original repository.

Changes

Rather than using AB3DMOT's max_age and min_hits parameters, this Gaussian mixture filter uses more conventional parameters for managing Gaussian mixtures (RFS Example), namely:

• Each Gaussian component is weighted (w_i, [0,1]), weight is directly proportional to classification score and evolves according to it.

• Probability of survival (p_s [0,1]) - probability of target survival between a consecutive time steps.

• Estimate threshold (t_e [0,1]) - threshold above which to report tracked targets.

• Prune threshold (t_p [0,1]) - threshold below which to remove Gaussian components from the tracker.

• Small part of Mahalanobis distance code is taken from Probabilistic 3D Multi-Object Tracking for Autonomous Driving.

• A model that is closer to the constant velocity model (even not exactly) is used for tracking (delta timestep is taken into account).

Essentially, only probability of survival p_s and estimate threshold needs to be tuned as prune threshold can be set to some low value according to performance considerations.

Results on Argoverse Leaderboard

Leaderboard classes: (C)ar, (P)edestrian.

	C:MOTA	P:MOTA	C:MOTPD	P:MOTPD	C:MOTPO	P:MOTPO	C:MOTPI	P:MOTPI	C:IDF1	P:IDF1
Simple GMF	71.54	49.62	0.33	0.36	11.60	23.20	0.18	0.18	0.81	0.60
Baseline	65.90	48.31	0.34	0.37	15.97	25.04	0.20	0.18	0.79	0.58

Running the Code

TODO

Citing this work

Open-source Implementation

@misc{ author = {Andrey Pak}, title = {Simple Gaussian Mixture Filter (Tracker)}, howpublished={\url{https://github.com/apak-00/argoverse_simple_gmf_tracker/}}, year = {2020}, }