Keywords: classification, signal processing, statistical modelling
Wearable accelerometers provide an objective measure of human physical activity. They record high frequency unlabeled three-dimensional time series data. We extract meaningful features from the raw accelerometry data and based on them develop and evaluate a classification method for the detection of walking and its subclasses, i.e. level walking, descending stairs and ascending stairs. Our methodology is tested on a sample of 32 middle-aged subjects for whom we extracted features based on the Fourier and wavelet transforms. We build subject-specific and group-level classification models utilizing a tree-based methodology. We evaluate the effects of sensor location and tuning parameters on the classification accuracy of the tree models. In the group-level classification setting, we propose a robust feature inter-subject normalization and evaluate its performance compared to unnormalized data. The classification accuracy on the subject-specific level was on average 87.6%, with the ankle-worn accelerometers showing the best performance with an average accuracy 90.5%. At the group-level, the average classification accuracy using the normalized features was 80.2% compared to 72.3% for the unnormalized features. In summary, a framework is provided for better use and feature extraction from the raw accelerometry data to differentiate among different walking modalities as well as considerations for study design.