why would you need machine learning? just detect a minimum of one bright light source, assume it's a headlight, and switch the lights. What the worse a false positive could do? this doesn't need AI
Comparable systems based on inductive loops work during the day. A design based on detecting headlights wouldn't. Software/compute is cheap and as you said reliability is not a safety concern.
We’ve had camera-triggered intersections for years now in Estonia. They seem to work equally well in daylight as well as darkness. I don’t think you need any ML for it either, you basically only need to detect motion within a predefined region which can be done with traditional CV algos quite reliably.
