There really aren't any good trick to it-- it's just a matter of finding a good vantage point and watching each driver go by. You're supposed to note the color of the belt (black) and the color of the driver's shirt before pulling the car over. As you'd imagine, fatigue comes quickly-- I never did more than an hour or two at a time. (I still have no idea how the highway patrollmen can it all day, every day without losing their minds.)

NHTSA provides substantial grant funding for traffic-safety enforcement, conditional on the department's enforcement statistics for a particular set of safety-oriented violations, especially seatbelts. The brass encouraged us to take a minimum number of enforcement actions to maximize the chances of having the grant expanded. (It paid a for substantial chunk of our salaries and offered regular 1 1/2 overtime.)

The better way to make people wear their seatbelts would be to standardize a really annoying seatbelt alarm in every car. I have a Subaru Forrester, and if I unbuckle my belt for 10s in my driveway, it goes crazy.

Fun fact: At the police department, we actually had the seatbelt alarms for all of our cruisers disconnected. Being able to get quickly in and out without the seatbelts snagging on our gear was the priority at low speed. Most officers are trained to unbuckle and be ready to exit the vehicle before coming to a full stop (technically illegal.)

The truly cool application of technology to traffic enforcement is the speed camera lottery in Stockholm. I have very mixed feelings on automatic license plate readers: rules for what can and can't be done with the data they produce don't really exist, and that's a real problem.

I think it would be a very hard problem. Take a look at some pictures of cars in traffic-- the seatbelt is extremely hard to see. If it were doable, it would probably take more optics engineers than ML scientists; glare and lack of contrast look like the biggest problems.