Basically it looks at price deltas, (which is like first derivative) and smooths those out over period to account for noise.
But that just gives us velocity. What we want is accelleration (second derivative). So it takes the deltas of the smoothed deltas and also smooths it. Yes it lags a bit more as a result, but it's much cleaner and not affected by noise.
The motivation is to detect changes in accelleration as opposed to velocity. i.e. we could still be going up but someone is slamming the brakes and we don't stop right away, we detect brakes being applied.
This could be used as additional confirmation to your already existing system. Where it fails is in a ranging market you need some filter or additional signal to filter those out.