Recurrence Quantification Analysis for Human Activity Recognition Thang Ngo, Student Member, IEEE, Benjamin T. Champion, Matthew A. Joordens, Member, IEEE, Andrew Price, David Morton, Pubudu N. Pathirana, Senior Member, IEEE, Abstract— Human Activity Recognition (HAR) is a central unit to understand and predict human behavior. HAR has been used to estimate the levels of a sedentary, monitor lifestyle habits, track the levels of people’s health, or build a recom- mendation system. Many researchers have utilized the inertial measurement unit as an input tool to explore the HAR land. The recurrence plot (RP) technique recently has its applications diverse in various areas. From the recurrence plot, a machine- auto or hand-crafted approach can be used to extract feature vectors. While the machine-auto based approach has been reported in the literature, the latter hand-crafted based method has not. For that reason, this paper evaluated and demonstrated the feasibility of utilizing Recurrence Quantification Analysis (RQA), which was a typical hand-crafted method from RP, to classify human activities. A Linear Discriminant Analysis classifier yielded a 95.08% accuracy, which belonged in the top accuracy reported in the literature. Compare to the machine- auto or end-to-end approach, RQA is a far less complicated and more lean system that should be further analyzed in a HAR application. I. INTRODUCTION Human activity recognition (HAR) system roots its appli- cations in many fields. For example, medical applications of HAR included measuring immobility, monitoring drinking [1], and assisting the living of the elderly [2], [3], [4]. In military and security, HAR was a core module of surveillance cameras systems [5]. In industrial fields, HAR was used to track and provide assembly step guidance for maintenance workers [6]. Other applications included those of sports activities such as measurement and simulation of golf ball rotation, evaluation of golf swing skill, designing individual tailored gold clubs [7], [8], or human-computer interaction [9]. From the input source perspective, the HAR systems can be categorized into external and internal sources. The external system used inputs from camera, sensor stick while internal systems utilized inputs from wearable sensors. Such a camera-based system suits for massive surveillance appli- cation but contains several intrinsic shortcomings for per- sonal oriented systems. An example of such shortcoming is privacy as not everyone is comfortable with the idea of being recorded by a camera. Computational burden and pervasiveness are other shortcomings as it is not user-friendly or restricts subject to stay within a peripheral space. The internal system, such as motion inertial measurement unit Authors are with the School of Engineering, Deakin University, Waurn Ponds, VIC 3216, Australia (e-mails: tdngo@deakin.edu.au, benjamin.champion@deakin.edu.au, matthew.joordens@deakin.edu.au, dr.andrew.price@deakin.edu.au, david.morton@deakin.edu.au, and pubudu.pathirana@deakin.edu.au). (IMU) sensors, received an extensive development of mi- croelectronics during the past decade. Wearable sensors are now smaller have long-life batteries, more powerful, and are a part of daily life. They provide unprecedented opportunities to solve problems concerning the latter. A sensor-based HAR system generally consists of two main stages. Firstly, feature engineering extracts measure- ments from the data set and the second stage is building a classification model to categorise activities. Feature engineer- ing is considered the most critical step in the hand-crafted features approach, which directly affects the performance of machine learning models. Within this context, various techniques have been developed to extract features out of IMU sensor signal. Recurrence plots in recent times have been found in numerous research areas [10], [11], [12], [13]. From a recurrence plot, researchers can extract features by image processing with deep learning [14], [15], [16] or by Recurrence Quantification Analysis (RQA). While many researchers tackled the problem by image-based approaches, RQA in contrary, remained unexplored in the literature. Hence, in this paper we evaluated the RQA approach for HAR and discussed the utility of the feature selection scheme. II. MATERIAL A. Participants A total of 46 participants (21 women, 25 men, mean age 34.3 ± 9.79 years) were measured on campus at the Deakin University, Australia. All subjects did not have any limitations impacting their maneuvers in performing tests following the protocol. Height (166.7 ± 10.2 cm) and weight (67.8 ± 16.9 kilogram) were measured using a digital height weight scale. Personal demographic disclosure was collected through self-assessment reports such as age, sex, dominant limb, and favorite activities. Ethics approval for this study was granted by the ethical committee of the Human Research and Ethics Committee, Deakin Univer- sity, Australia (HREC Reference Number: STEC-02-2019- RALLAPALLE). Informed consent was obtained for all subjects enrolled in the experiments. B. Protocol and Experimental Apparatus During the test, five IMU sensors were attached to each participant by elastic belts to collect the motion data (see Fig. 2 for attachment positions). The protocol comprised five tests: 1) Stance Test: Subjects stood erect with their feet to- gether and hands hung by sides for 30 seconds. The 978-1-7281-1990-8/20/$31.00 ©2020 IEEE 4616 Authorized licensed use limited to: Carleton University. Downloaded on September 14,2020 at 08:32:47 UTC from IEEE Xplore. Restrictions apply.