Hand-crafted Features vs Residual Networks for Human Activities Recognition using Accelerometer

Inertial sensors combined with supervised machine learning techniques are largely employed for automatic Human Activity Recognition (HAR). Machine learning scientists made available to the community a plenty of labeled datasets that permit, especially in the recent years, to develop sophisticated techniques, such the ones based on deep learning. These techniques have recently become very popular because they are highly accurate. Nevertheless, some researchers still use the combination of traditional classifiers, such as SVM and k-NN, with handcrafted features or raw signals. The aim of this paper is to investigate the robustness of traditional classifiers combined with hand-crafted features compared with an end-to-end deep learning solution based on a Residual Network. Experiments on four public datasets are presented and discussed.