Unsupervised nonparametric method for gait analysis using a waist-worn inertial sensor

This paper describes a nonparametric approach for analyzing gait and identifying bilateral heel-strike events in data from an inertial measurement unit worn on the waist. The approach automatically adapts to variations in gait of the subjects by including a classifier that continuously evolves as it ''learns'' aspects of each individual's gait profile. The novel data-driven approach is shown to be capable of adapting to different gait profiles without any need for supervision. The approach has several stages. First, cadence episode is detected using Hidden Markov Model. Second, discrete wavelet transforms are applied to extract peak features from accelerometers and gyroscopes. Third, the feature dimensionality is reduced using principal component analysis. Fourth, Rapid Centroid Estimation (RCE) is used to cluster the peaks into 3 classes: (a) left heel-strike, (b) right heel-strike, and (c) artifacts that belongs to neither (a) nor (b). Finally, a Bayes filter is used, which takes into account prior detections, model predictions, and step timings at time segments of interest. Experimental results involving 15 participants suggest that the system is capable of detecting bilateral heel-strikes with greater than 97% accuracy.