Exemplar Selection Methods to Distinguish Human from Animal Footsteps

The ”class discovery” problem is the problem of learning a classifier from a mixture of unlabeled and labeled training data, under the constraint that labeled training data exist for only N-1 of the N target classes. The task of distinguishing human from animal footsteps can be framed as a class discovery problem. When humans travel alone, every footstep sound is caused by a human foot, therefore labeled training examples for the ”human” class are abundant. When humans travel with animals, their footsteps are interspersed and/or overlapped in time; without a tedious labeling effort, there are no goldstandard labels specifying which species created each of the footstep events. This paper will describe three different types of class discovery algorithm: the mixed-vs-unmixed classifier, the generative class discovery algorithm, and the class of algorithms sometimes called ”self training.” Experiments using the ARL/Mississippi multisensory personnel tracking database will be reported. Experimental results suggest that the mixed-vs-unmixed classifier gives the best performance in distinguishing mixed vs. unmixed test tokens (recordings containing humans alone vs. humans with animals), and that the self-training method shows promise for the task of learning to distinguish between the discrete footfall sounds of humans and animals.