A perhaps-stupid question: Does the software involved attempt any real classification of objects in its field of vision? Or is it a neural network - where all sorts of "interesting" things tend to happen for inputs which are a poor match for its training set?