I think this is quite similar to how music fingerprinting (Shazam, etc.) works, except it deals with multiple channels.

EDIT: Their paper confirms it: "FAST adapts a data mining algorithm, originally designed to identify similar audio clips within large databases"