Elsevier

Measurement

Volume 147, December 2019, 106810
Measurement

Characterization of a wearable system for automatic supervision of fitness exercises

https://doi.org/10.1016/j.measurement.2019.07.038Get rights and content

Highlights

  • A machine learning solution for tracking and quantification of physical activity.

  • The use of non-invasive wearables based on low-cost Inertial Measurement Unit (IMU)

  • An extensive measurement campaign for validating the procedure performance.

  • Prediction of computational resources required in an embedded implementation.

Abstract

It is widely known that physical activity is an effective tool for preventing several diseases. However, unsupervised training may lead to poor execution quality, resulting in ineffective training, or even injuries in worst cases. Automatic tracking and quantification of exercise efforts by means of wearables could be a way to monitor the execution correctness. As a positive side effect, these devices help in motivating people, increasing the quantity of physical exercises of users and thus improving health conditions as well. Unfortunately, despite the availability of some commercial devices, their performance and effectiveness are not documented. This work proposes a new solution that exploits machine learning (ML) techniques (in particular Linear Discriminant Analysis, LDA) for analyzing data coming from wearable Inertial Measurement Units (IMUs). Efforts have been done in reducing the computational requirements, in order to be compatible with constraints in perspective of embedded implementation. The experimental campaign carried out to measure the performance showed an average accuracy, recall and precision on the order of 97%, 93% and 90%, respectively.

Section snippets

Introduction and motivation

Being physically active on a regular basis has many positive effects on human health. A number of recent studies have investigated this connection. The causal relationship between physical activity and cardiovascular diseases, thromboembolic stroke, hypertension, type 2 diabetes mellitus, osteoporosis, obesity, and colon/breast cancer has been deeply analyzed in the last years [1]. Recently, the focus has shifted to the quantification of the dose-response connection between physical exercise

State of art and related studies

A survey in EU countries showed that only 31% of respondents reported sufficient physical activity [13]. An example is related to senior citizens, a part of the population that would greatly benefit from this kind of activity. A research found out that 87% of elderly subjects have health-related and/or lack-of-knowledge barriers [14]. As stated in the Introduction, a possible solution is the adoption of automated supervision of exercise training, made possible by the advent of low-cost

The proposed approach

The basic idea of the proposed solution is shown in Fig. 1. One or more wearable IMUs acquire the user’s movements while performing exercises. These data are then transmitted to a processing unit for data evaluation, for example through the broadly-used Bluetooth (BT) connection [33]. Following recent trends in personal systems for medical and sport/fitness support, the latter could be implemented with a smart device, such as a smartphone, tablet and smartwatch. Indeed, the term “mHealth” has

The actual implementation

The proposed system adopts a single non-invasive wrist-worn IMU, the x-IMU manufactured by x-IO technologies (http://x-io.co.uk). It hosts a tri-axial 12-bit accelerometer, a 16-bit tri-axial gyroscope and a 12-bit tri-axial magnetometer (not used in this work). The adopted sampling frequency is 256 Sa/s. An example of raw accelerometer and gyroscope data can be seen in Fig. 2.

The x-IMU sensor is equipped with a BT module and has an internal micro SD card for data storage. In the experimental

Experimental results

As previously explained, the feature reduction requires a preliminary PCA analysis in order to extract the coefficients for the linear mapping of the initial 72 features into the reduced 20 dimensional feature space. Moreover, the LDA must be trained using a reference data set. In the proposed approach, only a single reference data set has been selected for both the aforementioned tasks; in particular, the “leave-one-out” cross-validation (LOOCV) has been applied on the overall acquisitions of

Conclusions

In this paper, a new approach for automatic classification and counting of fitness exercises is proposed. In particular, PCA and LDA ML techniques have been employed in light of reducing the computational efforts. Additionally, a confidence parameter and a filtering strategy based on event-duration have been introduced, in order to improve reliability. The comprehensive performance characterization leverages on a data set obtained monitoring the physical activity of seven users with different

Funding

This work has been partially supported by Smart Cities and Communities and Social Innovation Grant D84G14000220008: “Smart Aging” and by the University of Brescia H&W grant “Breaking bad breakfast habits” and “Work Wealth Production Productivity”.

Declaration of Competing Interest

None.

References (47)

  • A. Depari et al.

    Multi-sensor system with Bluetooth connectivity for non-invasive measurements of human body physical parameters

    Sens. Actuators, A

    (2013)
  • T. Fawcett

    An introduction to ROC analysis

    Pattern Recogn. Lett.

    (2006)
  • G. O’Donovan et al.

    The ABC of Physical Activity for Health: a consensus statement from the British Association of Sport and Exercise Sciences

    J. Sports Sci.

    (2010)
  • M. Hamer et al.

    Walking and primary prevention: a meta-analysis of prospective cohort studies

    Br. J. Sports Med.

    (2008)
  • W.L. Haskell et al.

    Physical Activity and Public Health: Updated Recommendation for Adults from the American College of Sports Medicine and the American Heart Association

    Off. J. Am. Coll. Sports Med.

    (2007)
  • A. Deslandes et al.

    Exercise and mental health: many reasons to move

    Neuropsychobiology

    (2009)
  • M. Hamer et al.

    Dose-response relationship between physical activity and mental health: the Scottish Health Survey

    Br. J. Sports Med.

    (2009)
  • M.D. Bravata et al.

    Using pedometers to increase physical activity and improve health

    J. Am. Med. Assoc.

    (2007)
  • V. Andalibi et al.

    Data correction for seven activity trackers based on regression models

    37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), August 2015

    (2015)
  • D. Steffen et al.

    A personalized exercise trainer for elderly

    5th International conference on Pervasive Computing Technologies for Healthcare (PervasiveHealth), May 2011

    (2011)
  • M. Sjöström et al.

    Health enhancing physical activity across European Union countries: the Eurobarometer study

    J. Public Health

    (2006)
  • H. Koskimäki et al.

    Recognizing gym exercises using acceleration data from wearable sensors

  • P. Siirtola et al.

    Efficient accelerometer-based swimming exercise tracking

  • Cited by (0)

    View full text