Classification of waist-acceleration signals in a continuous walking record

We attempted to distinguish walking on level ground from walking on a stairway using waist acceleration signals. A triaxial accelerometer was fixed to the subject's waist and the three acceleration signals were recorded by a portable data logger at a sampling rate of 256 Hz. Twenty healthy male subjects were asked to walk through a corridor and up and down a stairway as a single sequence, without any instruction. The data were analyzed using discrete wavelet transform. Walking patterns were classified in two stages. In the first stage, the times when the walking pattern changed were detected using the low-frequency component of the anteroposterior acceleration (LF(A)) and of the vertical acceleration (LF(V)). In the second stage, the three types of walking patterns were classified by comparing powers of wavelet coefficients in the vertical direction (P(WCV)) and in the anteroposterior direction (RP(WCA)). Changes in walking patterns could be detected by using either LF(A) or LF(V). Walking down stairs could be distinguished from the other types of walking as it gave the largest value in P(WCV), and walking up stairs could be discriminated from level walking using RP(WCA). Level and stairway walking could be classified from continuous records of waist acceleration.