Problem about filtering results changes while updating gait status real-time

[ZM-NEU]

Recently while trying to modify this algorithm for real-time application I ran into a problem: Firstly I tried to update the result every 0.5 second or so to detect gait phases(stance or swing) but I found that even with the same dataset, Butterworth filtered acceleration magnitudes can be quite different when calculated as a whole sequence and calculated after been divided into several sequences. That resulted in quite serious inaccuracy while these acceleration magnitudes are used to detect gait phases, not only the start and end of each step, but also the number of steps taken are different. The "stationary or not" result would be at least 11% percent different. Does anyone have ideas about how to deal with such problem? Thanks in advance!