I was actually spending quite some thought about this. At least for water parks, I think this should be doable. My idea was to have a wide angled camera above the pool (distortion is not a problem here), with an algorithm that would recognize any structure that is not moving for a given time (lets say 30 seconds). This would create many false positives, probably, but the system could just activate a buzzer and zoom to the spot on a display that would be permanently monitored by some lifeguard. The number of false positives could probably be dramatically reduced, if there was some way to only react to non-moving-structures underneath the surface, maybe by some cameras from the side?
Problem is, when I was doing my lifeguard training they showed us / told us how the surface of the water distorted swimmers and who / what is below the surface. A towel or swim toy can look like a person once fully submerged. (it is required to wear polarized sunglasses in chair) You would need cameras under the water, focused on the bottom of the pool to be of any use I imagine. As a lifeguard, our main objective is scanning for signs of struggle yes, but also to be checking for bodies below the surface. With unresponsive victims, every second counts.
