Rich learning representations for human activity recognition: How to empower deep feature learning for biological time series.

Abstract

Deep learning versus feature engineering has drawn significant attention specifically for applications where expertly crafted features have been used for decades. Human activity recognition is no exception where statistical and motion specific features have shown potential in detecting falls and other daily activities across a wide range of datasets. This paper provides an in-depth study and comparison of two fundamentally different approaches to HAR while introducing a novel way to harness the spectral properties of biological time series in addition to temporal features. A research group at Stanford recently proposed subject agnostic features as state-of-the-art when applied to a large dataset with many participants of different ages. In this paper, we demonstrate that implicit feature learning in the latent spaces of deep learning algorithms can be powerful alternatives to using finely tuned domain-specific features for HAR. In fact, when using a spectrotemporal representation of the raw sensor data in the form of spectrograms, a standard convolutional neural network without any prior conditioning on the features, statistically significantly outperforms the prior state-of-the-art using subject agnostic features in all the different partitions of the dataset with a significant  29.8% reduction in the overall average error rate.