It's generally presumed that a background (either an input sample or non-enriched regions) will have signal following a Poisson distribution. Consequently, one can track the lambda from that (see the --llocal and --slocal options) see if a given spot has siginificantly more signal than expected given that distribution. This is then given a p-value accordingly and you have your peak.

Edit: The wiki from MACS2 may be helpful.