A Survey of Sensor Modalities for Human Activity Recognition

Bruce Yu, Yan Liu, Keith Chan

Abstract

Human Activity Recognition (HAR) has been attempted by various sensor modalities like vision sensors, ambient sensors, and wearable sensors. These heterogeneous sensors are usually used independently to conduct HAR. However, there are few comprehensive studies in the previous literature that investigate the HAR capability of various sensors and examine the gap between the existing HAR methods and their potential application domains. To fill in such a research gap, this survey unfastens the motivation behind HAR and compares the capability of various sensors for HAR by presenting their corresponding datasets and main algorithmic status. To do so, we first introduce HAR sensors from three categories: vision, ambient and wearable by elaborating their available tools and representative benchmark datasets. Then we analyze the HAR capability of various sensors regarding the levels of activities that we defined for indicating the activity complexity or resolution. With a comprehensive understanding of the different sensors, we review HAR algorithms from perspectives of single modal to multimodal methods. According to the investigated algorithms, we direct the future research on multimodal HAR solutions. This survey provides a panorama view of HAR sensors, human activity characteristics and HAR algorithms, which will serve as a source of references for developing sensor-based HAR systems and applications.

Download


Paper Citation