Just as I was reading your comment, a boat went by or something and there was a sudden rush of sand and debris kicked up from the bottom of the feed, and for the past 5 minutes or so, there's been a constant stream of debris flying by and swirling around. Looks more like TV static than "still and opaque". Sometimes larger pieces of debris fly by.
But that said, nothing here that proper machine learning couldn't handle.
I think it goes both ways though. One can justify more complex solutions, only to not need the added complexity. One can always ignore edge cases to justify a more simple solution. only to find out that the cut functionality was not quite as "edge" as expected.
My security cameras love to declare that they detected motion because a car drove by and swept its headlights across my porch, and my cameras use a simple change detector.
What are the consequences of false positives? For me, I get a notification, check it, and roll my eyes. For a lock, it opens too often? I have no idea what the consequences of that are. Maybe that's perfectly acceptable and a simple change detector is enough. Or maybe it changes water balance somehow and causes ecological problems, or impacts boat navigation (I mean, the lock is there for a reason in the first place right?)
Sometimes complex solutions are justified. A camera is more complex than just looking into the water with your naked eye. But that necessitates having someone physically present, and able to see through the water. Everyone agrees that the complexity is worth it.
I've worked with too many PMs who want the fastest easy solution without regard for the actual use case and then get perplexed by the side effects (that they were previously told about but refused to grasp). And generally I wind up being the person forced to be physically present watching for fish because the PM thought adding a camera was too much complexity.
But that said, nothing here that proper machine learning couldn't handle.