My guess from the NTSB description is that their tracker had object category as part of the track information, and that detections of one category weren't allowed to match a track of another category. This is useful if, say, you are tracking a bicycle and a pedestrian next to each other. When you go to reconcile detections with tracks, you know that a bicycle isn't going to suddenly turn into a pedestrian, this rule helps to keep you from switching the tracks of the two objects.

Unfortunately that also means that when you have a person walking a bicycle that might be recognized as either one, this situation happens, and they initialize the velocity of a newly tracked object as zero.

Even this though shouldn't have been enough to sink them, because the track should have had a mean-zero velocity, but with high uncertainty. If the uncertainty were carried forward into the planner, it would have noted a fairly high collision probability no matter what lane the vehicle was in, and the only solution would be to hit the brakes or disconnect.

Furthermore, if you've initiated a track on a bicycle, and there's no nearby occlusion, and suddenly the bicycle disappears, this should be cause for concern in the car and lead to a disconnect, because bicycles don't just suddenly cease to exist. They can go out of frame or go behind something but they don't just disappear.