Descriptors for Human Activity Recognition

Abstract

This article presents some filtration process on a public human activity dataset called `EJUST-ADL-l`. It consists of four types of 3D IMU sensory signals: User acceleration, angular velocity, rotation displacement, and gravity for 14 activities of daily living ADLs measured by a wearable smart watch. The EJUST-ADL-l dataset contains mainly activities of communication, feeding, transferring, and personal grooming. The data is filtered by using several descriptors. The descriptors are constructed using different combinations of the following signal features: The minimum, maximum, median, mode, range (maximum-minimum), mean, standard deviation, entropy, and the autocorrelation function up to a certain lag and taking these values as representative features of the given signal. Experiments show which descriptor achieves highest accuracy on the random-forest based model.