I'm shelving this project for now, because it's been successful enough and any improvement would take a lot of time and effort.  I've gotten the AI to work on both humans and squirrels (but only have video of humans, which I'm not going to post here).  For now, I'm just using YOLOv3 running on my laptop to do the heavy lifting, and it just uses the webpage to fire.  It's 100% sketchy and bodgey, but it works.  I've tried to develop a CNN that can do it onboard the pi to remove the need for an internet connection, but it turns out that's hard.  I don't have much experience with tensorflow lite, and couldn't figure out the installation on the pi zero w.  Additionally, the zero w only has one core, so running darknet yolo with the NNPACK option won't improve anything, unlike on the full-sized PIs.  

The one clear solution is more computing power. I think that the easiest way to do this would be the Intel Modivius NCS.  Others have used this to do amazing things with the same pi zero w, so I think it'll work.  It would somewhat compromise on the aesthetics of the project, and would make me more hesitant to leave the gun alone outside somewhere, but it wouldn't be too bad because it can just plug into the existing USB port on the side of the gun.  Anyways, that is a pretty significant monetary and time investment, and I'm happy enough with this project as it is now.