A Localized Learning Approach Applied to Human Activity Recognition

The recognition of human physical activities and postures based on sensor data has received much research attention in several human health and biomedical engineering applications. In this article, the challenges of class imbalance and ambiguity (or confusion) are discussed that frequently arise in data from human activity recognition (HAR) systems. In order to reduce the influence of imbalance and ambiguity in HAR problems, a novel hybrid localized learning approach of K-nearest neighbors least-squares support vector machine is proposed. The classifier is applied to different synthetic and real-world datasets where imbalance and ambiguity are present. In this article, it is novel to apply a hybrid localized learning algorithm to the HAR problem. When compared to different global and local approaches, higher classification performances could be obtained by using the proposed localized learning approach. Furthermore, the computational effort could be reduced in an online learning mode.