Advanced classification of ambulatory activities using spectral density distances and heart rate
Introduction
Physical activity recognition (PAR) is a topic extensively researched nowadays in the biomedical domain. As a health status indicator, PAR is capable of providing feedback about the individual behavior, and is thus used in the treatment of adult overweight and obesity as well as the supervision of the balance and the locomotion of the elders and the disabled in order to avoid the physical risks they might be subjected to, falling for example [1], [2]. Furthermore, physical activity (PA) information is also the key to home-based rehabilitation and physical therapies for some pathology [3], [4]. Hence, monitoring the daily life activities and identifying them automatically become the research focus in these fields. From the application point of view, it is essential to pair the recognition precision with the basic demands like patient confidentiality, accessibility, user-friendliness and affordability [4].
The present study has a two folded contributions. Firstly, it proposes a novel method for activity classification using the spectral density distance measures from acceleration data. Secondly, it studies the influence of combining the spectral density distance and HR information in the enhancement of the classification performances. The paper is organized as follows: Section 2 analyzes the state of the art in PAR, Section 3 presents the system architecture and the experimental protocol for data collection. Section 4 explains the interest of the distance measure using spectral density and Section 5 gives methodological details on the AR spectral model training and the inclusion of HR for activity classification. Section 6 compares the proposed method with the reference feature selection method in the literature [12] and outlines the most important contributions. Finally, Section 7 summarizes our study.
Section snippets
Related work
The traditional monitoring techniques were either based on physical observations: surveys, questionnaires and self-reports or on pedometers and actimeters. Thanks to the revolution of sensor technologies, advances in miniaturizing the size, the weight and the cost of commercially available inertial sensors, accelerometers become the most popular choice for PAR [5]. The raw output of accelerometers is processed and transformed into pertinent information: the PA type, as well as its intensity,
Data collection
The dataset comprised recordings from eight volunteers. The participants were recruited from University of Rennes 1 and from the Ecole Normale Supérieure (ENS) − Rennes, France, from both sexes, aged between 18 and 30 years, healthy with different levels of physical fitness (Table 1). The data were collected at the Ker Lann stadium, Bruz.
Each subject was asked to perform five ambulatory activities in a random order: running, walking, cycling, car riding (as a passenger) and resting. Three
Limitation of classical approaches
It is difficult to recognize the pattern for each PA using the raw acceleration signals due to the fact that they are by nature noisy and containing repetitive variations. In general, typical procedures of a PAR system start by extracting basic statistical features from the signals in the time and frequency domain, then reduce these feature dimensions in order to choose the most relevant features to discriminate PA, and finally recognize the PA pattern using a classification tool [11], [12].
Set of spectral distances features
As explained in [22], the SpD method relies on the fact that power spectral densities of a random process (activity’s signal) include information of the periodicity and the cyclostationarity of the process. Hence, spectral distances can be used to measure the differences between two such processes [24].
The Itakura-Saito distance dIS is the spectral distance considered in our study. It is based on the power spectral density (PSD) of a random process obtained by the DFT (Discrete Fourier
Evaluation
In this section, we discuss and evaluate the system on the basis of three comparative studies. The first one compares our proposed method with the state-of-the-art method in [12]. The second one compares the responses of the system regarding the sensors’ placements. The third one measures the impact of the HR information inclusion on the system accuracy.
To test the strength of the system regarding the classification algorithms, we have chosen to apply three different types of classifiers: the
Conclusion
In this paper, we developed a classification method based on spectral distance measure, the overall mean accuracy proves the pertinence of this technique to automatically distinguish between locomotion behaviors while reducing the computation and the complexity of the classification algorithms. We believe that the spectral features contain the sufficient information needed to classify PA. A new reduced set of spectral-based features (60 spectral distances when considering all the sensor
Acknowledgments
This work, part of the SHERPAM project (2014–2018), has received a French government support granted to the CominLabs excellence laboratory, which is managed by the National Research Agency in the “Investing for the Future” program under reference ANR-10-LABX-07-01.
Further information about this project can be found at http://www.sherpam.cominlabs.ueb.eu/.
References (25)
- et al.
Online classifier construction algorithm for human activity detection using a tri-axial accelerometer
Appl. Math. Comput.
(2008) - et al.
Comparative study on classifying human activities with miniature inertial and magnetic sensors
Pattern Recogn.
(2010) - et al.
Detection of posture and motion by accelerometry: a validation study in ambulatory monitoring
Comput. Hum. Behav.
(1999) - et al.
Centinela a human activity recognition system based on acceleration and vital sign data
Pervasive Mobile Comput.
(2012) Distance measures for signal processing and pattern recognition
Signal Process.
(1989)- et al.
Mobility assessment in older people: new possibilities and challenges
Eur. J. Ageing
(2007) - F. Hijaz, N. Afzal, T. Ahmad, O. Hasan, Survey of fall detection and daily activity monitoring techniques (IEEE, 2010,...
- et al.
Clinical applications of sensors for human posture and movement analysis: a review
Prosthet. Orthot. Int.
(2007) - et al.
Best practices for using physical activity monitors in population-based research
Med. Sci. Sports Exerc.
(2012) - et al.
Evolution of accelerometer methods for physical activity research
Br. J. Sports Med.
(2014)
A survey on human activity recognition using wearable sensors
Commun. Surv. Tutorials IEEE
Machine learning methods for classifying human physical activity from on-body accelerometers
Sensors
Cited by (5)
A novel algorithm for minute ventilation estimation in remote health monitoring with magnetometer plethysmography
2021, Computers in Biology and MedicineCitation Excerpt :This study is part of the Sherpam (Sensors for Health Recording and Physical Activity Monitoring) project (https://project.inria.fr/sherpam/) that aims to develop a new generation monitoring system to allow health monitoring at home, during indoor and outdoor activities (at work, shopping, sports practicing, etc). Concretely, the project has the ambition to improve the recognition of PA [22] and the estimation of associated energy expenditure in healthy subjects [23,24]. A modular Android application aggregates data from various types of sensors and transmits these data to a remote aggregation server for storage and analysis.
An ECG classification based on modified local binary patterns: a novel approach
2021, Research on Biomedical EngineeringComparison of offline and real-time human activity recognition results using machine learning techniques
2020, Neural Computing and ApplicationsRecognition rate difference between real-time and offline human activity recognition
2017, Internet of Things for the Global Community, IoTGC 2017 - ProceedingsDeleterious effect of participant positioning on the acceptability and acceptance of a wellness management system under development
2021, Applied Sciences (Switzerland)