It's surprising that people don't rely more on automatic arrival time picking + simple outlier rejection. It's so much more efficient than these waveform correlation approaches while being much simpler.
The FAST article mentions how using the ratio of short term seismic activity to long term activity to detect seismic events is prone to false positives. With this approach, you "pick" a discrete time when the seismic wave first reaches a given station, and then using knowledge of the speed of sound in the subsurface you perform nonlinear regression to locate the the point in time and space that minimizes the travel time error. However, this gets tricky when your algorithm picks the wrong arrival time, and picks a truck driving by instead of the actual seismic wave. Now your estimate of when and where the earthquake occurred has a substantial amount of error.
If you have enough stations for the seismic detection to be over defined (n > 4), one way around this is to consider all spikes in seismic activity as event picks, and then iteratively discard outlier picks until you're either left with an earthquake detection that has a residual travel time error within some acceptance threshold, or you run out of picks to make an over defined solution. This approach ends up being very, very fast.
EDIT: Their paper confirms it: "FAST adapts a data mining algorithm, originally designed to identify similar audio clips within large databases"