I think it's entirely possible to determine with high confidence that someone is wearing a seat belt thru video analytics.
Even in that case, the analytic may have low false positives (i.e., identifying someone as wearing a seatbelt who is not). However, because of obstructed video frequently, it would still miss lots of people wearing a seat belt.
That said, I agree with your point about the implementation being about focusing on those NOT wearing a seat belt.
For people who hypothetically what to know about who wears vs does not wear seat belts, is this for statistics / trend analysis or real-time alerting / intervention? The statistics are going to be way way off because of the issues described above. You might get some correct alerts but is the goal to mail them tickets or to dispatch a police officer? I am genuinely curious if there is a real demand for this outside of it being something neat to do.