That's only true if you want hardware that sometimes detects some kind of vehicles, in the right weather and from the right angles.
Truly reliably detecting this is much further away. The consequences of not getting it right are not as catastrophic as getting it wrong in a self-driving car, but they could still mess traffic up pretty badly.
Your reaction might be that it has been $50 for like 5 or 10 years now, and if so my response is that in this area, that is recent. It's a pretty slow-moving, late adopter industry. Most of the equipment getting installed probably hasn't been redesigned in years or decades. The people approving stuff want to buy from known vendors. The vendors are making money steadily now and don't want to go out on a limb with unproven tech.
Which doesn't mean it's impossible. I interpret it the opposite way: the fact that it hasn't happened yet doesn't mean it won't or can't. I would expect it to take this long or longer.
Also, I think the angle for getting this adopted isn't really efficiency, it's cost saving. Right now, some lights are on sensors (instead of just timing) but they do it with induction loops (https://en.wikipedia.org/wiki/Induction_loop). If you could match or surpass the reliability of those at a lower cost, you might get people to switch.
But once you've crossed that hurdle, you can improve efficiency too, of course.
Of course, you have to keep in mind that the camera has to be mounted in the proper position. So it's probably not really $50 of hardware unless you can (in most cases, at least) piggyback on something that already exists. Although that seems possible because the signal light does need to be visible by cars approaching the intersection.
$50 * 300 000 (quick google search about the number of traffic light in the US), That's not taking into account installation and regular maintenance.
I'm all for optimising things but sometimes it's just not worth the hassle. A well timed orange light is good enough.
This bugs me. $50 of hardware can accomplish this, today. Raspi running vehicle detection code. If something is coming and the other way is clear, why not flip the intersection before it arrives?