How to better track cars using trackerGNN using constraints

Using a Lidar Sensor on a vehicle, GPS and SLAM I detect surrounding cars in 0.1 Hz increments. Using trackerGNN I want to get the trajectories of these cars. In busy and noisy scenes however, the tracker gets confused. Sometimes a car is driving in one direction and a car on the opposing lane (that is yet undetected) obscures it for a moment and is then detected, the tracker thinks it's the same car. Sometimes noise can create trajectories with insane speed and acceleration. Since cars are slow moving, don’t change direction immediately and, in my case, don’t drive backwards, is there a way to implement these constraints in trackerGNN (or any other tracker)?

Now obviously tracking can be improved by tuning the parameters better or even by filtering, merging and changing the trajectories after the tracking, but I am looking for a way to get these constraints directly into a tracker.