I couldn't find a paper that talks about the specifics of this very problem. The ones I've found in the past fail with outliers like a bike going through stopped traffic and or the scenario above.

The best approach I could think of is they expand the time they're looking at to classify the vehicles before the outlier situations.