The cross correlation idea may work well, but you will have to perform it with varying levels of stretching. 

The cross correlation will be computed using a 2D window of length L and height H representing the activity. Choose a moderate length L that resembles the expected length of the activity.  

The cross correlation will be taken between the signal and this window for *n* iterations where in each iteration the enlarging/shrinking factor takes a new value. You probably won't need the three channels.