Usman,
The GNN framework is considered the most simple and least robust multi-object tracking framework. It makes hard/crisp decisions, which are very useful in the case where you have few false alarms and the association is not ambiguous. In its basic form, trackerGNN works with a history-based track maintenance, in which decisions to confirm or delete a track are made based on how many detections were assigned to a track and how many steps the track went unassigned, respectively. This is another crisp (yes/no) kind of decision.
You can try to use the score-based track maintenance, which will also allow you to manage the log likelihood (score) of a track and decide when to confirm or delete it based on the score. While it does not solve the crisp assignment decisions, it can make the management a little more robust. You'll notice a few more properties are enabled when choosing the score management option. Use them to specify things like false alarm rate and birth rate (beta) in a sensor volume.
However, once there is high clutter, ambiguity due to uncertain measurements, etc., the association decisions themselves need to be more robust. This is where GNN fails, unforunately.
That's why JPDA is considered a great alternative for a fast yet more robust association algorithm. If you used the trackerGNN, you can switch to the trackerJPDA with relatively ease. It takes the same inputs and provides the same outputs. It also shares some of the basic properties with trackerGNN. If you set the track management in JPDA to "integral" the algorithm switches to JIPDA, which uses an integrated track probability in association, management, and udpate. This is the "best" JPDA for a case of high clutter and highly ambiguous associations.
A level up from there would be using trackerTOMHT, which can entertain multiple association hypotheses. Each association hypothesis will branch the tracks. This is computationally expensive, but considered the best at managing ambiguous cases. You can use trackerTOMHT in the same way you used trackerGNN / trackerJPDA.
Consider this example as a guideline.
https://www.mathworks.com/help/fusion/ug/tracking-closely-spaced-targets-under-ambiguity.html
