A dynamically detected bounding box (cat) overlapping with a predefined bounding box (bowl) would be less than 50 lines of python if you use openCV. Discriminating raccoons and cats would probably be another 50 lines if you use some pre-trained ML model.

Adding an electric scale to the bowl would certainly cut down the false positive rate you're going to get (from the cat, safe to say the raccoon will always be eating if it overlaps the bowl).