Information Augmentation for Human Activity Recognition and Fall Detection using Empirical Mode Decomposition on Smartphone Data

In this paper, we propose a novel design to reduce the number of sensors used in activity recognition and fall detection by using empirical mode decomposition (EMD) along with gravity filtering so as to untangle the useful information gathered from a single sensor, i.e. accelerometer. We focus on reducing the number of sensors utilized by augmenting the information obtained from accelerometer only given that the accelerometer is the most common and easy to access sensor on smartphones. To do so, one gravity component and three intrinsic mode functions (IMFs) are extracted from the accelerometer signal. In order to assess how informative each component is, the raw components are directly used for classification, i.e. without hand-crafting statistical features. The extracted signal components are then individually fed into parallelized random forest (RF) classifiers. The proposed design is evaluated on the publicly available MobiAct dataset. The results show that by only using accelerometer data within the proposed scheme, it is possible to reach the performance of two sensors (accelerometer and gyroscope) used in a conventional manner. This study provides an efficient and convenient-to-use solution for the smartphone applications in human activity recognition domain.