I'm impressed that he can get low false positives with such simple methods. That's great to know.
If he tried using line detection, vectorize it, then train the algorithm with the tagged shapes of squirrels. Would that be too slow to do in real time? Or could it work to reduce false positives?
Adding complementary features such as edges (I'm guessing this is what you mean by lines) tends to improve the accuracy. It would also be possible to do this in real time.
I don't have time to watch the full video so I don't know what features he is currently using, but in object detection nowadays most people are using some variant of the SIFT descriptor. These are built not upon edges but on the image gradient per-pixel. The current 'hot' feature in terms of frequency of use is probably Histograms of Oriented Gradients (HOG) which do exactly what they say on the tin: Take a region of the image and count how many times a particular gradient direction occurs and the total magnitudes. Slightly more difficult to run in real-time but libraries exist.
I didn't show the false positive video. :P There were some additional features thrown in, but for presentational simplicity I went with some strong, easy ones to explain. I think one thing helped keep FPs low was some frame-based signal thresholding. So, I required a significant squirrel signal (over multiple frames) before it went hot. That prevented any small FP blip to trigger the gun.
If he tried using line detection, vectorize it, then train the algorithm with the tagged shapes of squirrels. Would that be too slow to do in real time? Or could it work to reduce false positives?