I agree, there needs to be multiple orders of magnitude improvements in HW&SW, as well as domain specific dataset development for this to work at all.
I can imagine using a very accurate image segmentation algo to analyze the background and thereby prevent shooting lasers at vulnerable targets, but it's still hard to see how this is a good idea.
A 1W laser is even more dangerous than you say. It is powerful enough to start wood fires.
An interesting future application for computer vision, but until both hardware and software increase in speed and precision by several orders of magnitude, this is not feasible.
Current neural object detection just isn't there yet.
Not to mention the dangers of shooting a powerful 1W laser out in the open...
A 1W laser is even more dangerous than you say. It is powerful enough to start wood fires.