If you combine the two properly to get true orientation information you'll be able
do appropriate gesture recognition (which is what this is really) using the angular
data. You do need a proper DCM or quarternion orientation library to get this,
as used in IMUs. That basically uses the gyro only (except the accelerometer corrects
for drift).
However you can take the higher frequency information from the accelerometer
(the spikes) to indicate foot fall I suspect. This would be a second technique to
back up the first.
One of the great signal processing techniques you might want to play with is
correlation, its often imployed to recognise signals buried in noise and interference.