Quantifiable fitness tracking using wearable devices

Abstract

Monitoring health and fitness is emerging as an important benefit that smartphone users could expect from their mobile devices today. Rule of thumb calorie tracking and recommendation based on selective activity monitoring is widely available today, as both on-device and server based solutions. What is surprisingly not available to the users is a simple application geared towards quantitative fitness tracking. Such an application potentially can be a direct indicator of one's cardio-vascular performance and associated long term health risks. Since wearable devices with various inbuilt sensors like accelerometer, gyroscope, SPO2 and heart rate are increasingly becoming available, it is vital that the enormous data coming from these sensors be used to perform analytics to uncover hidden health and fitness associated facts. A continuous estimation of fitness level employing these wearable devices can potentially help users in setting personalized short and long-term exercise goals leading to positive impact on one's overall health. The present work describes a step in this direction. This work involves an unobtrusive method to track an individual's physical activity seamlessly, estimate calorie consumption during a day by mapping the activity to the calories spent and assess fitness level using heart rate data from wearable sensors. We employ a heart rate based parameter called Endurance to quantitatively estimate cardio-respiratory fitness of a person. This opens up avenues for personalization and adaptiveness by dynamically using individual's personal fitness data towards building robust modeling based on analytical principles.

Data Fusion on Motion and Magnetic Sensors embedded on Mobile Devices for the Identification of Activities of Daily Living

“…The driving, walking, running, cycling, resting, and jogging was recognized by ANN with mean, minimum, maximum, standard deviation, difference between maximum and minimum, Parseval's Energy, Parseval's Energy in the frequency range 0 -2.5 Hz, Parseval's Energy in the frequencies greater than 2.5 Hz, RMS, kurtosis, correlation between axis, ratio of the maximum and minimum values in the FFT, skewness, difference between the maximum and minimum values in the FFT, median of troughs, median of peaks, number of troughs, number of peaks, average distance between two consecutive troughs, average distance between two consecutive peaks, indices of the 8 highest peaks after the application of the FFT, and ratio of the average values of peaks and troughs as features from the accelerometer and gyroscope data, reporting an accuracy between 57.53% to 97.58% [31].…”

“…The authors of [25] also used the mean and standard deviation as features for the application of KNN and SVM methods, in order to recognize walking, resting, running, going downstairs, and going upstairs with a reported accuracy higher than 90%.The standard deviation, maximum, minimum, correlation coefficients, interquartile range, mean, Dynamic time warping distance (DTW), Fast Fourier Transform (FFT) coefficients, and wavelet energy are extracted as features from accelerometer and gyroscope sensors, in order to recognize walking, jumping, running, going downstairs, and going upstairs with several methods, such as SVM, KNN, MLP, and Random Forest, reporting an accuracy between 84.97% and 90.65% [12]. The authors of [26] extracted the same features for the recognition of walking, going upstairs, going downstairs, jumping, and jogging activities, implementing KNN, Random Forests and SVM methods, reporting an accuracy of 95%.The authors of [27] extracted the variance, mean, minimum and maximum along the Y axis of the accelerometer, and the variance and mean along the X axis of the gyroscope, and implemented the SVM method for the recognition of running, walking, going downstairs, going upstairs, standing, cycling and sitting, which reports an accuracy of 96%.In [28], the authors extracted the skewness, mean, minimum, maximum, standard deviation, kurtosis, median, and interquartile range from the accelerometer and gyroscope data, implementing the MLP, SVM, Least Squares Method (LSM), and Naïve Bayes classifiers for the recognition of falling activities with a reported accuracy of 87.5%.The SVM, Random Forest, J48 decision tree, Naïve Bayes, MLP, Rpart, JRip, Bagging, and KNN were implemented in [29] for the recognition of going downstairs, going upstairs, lying, standing, and walking with the mean and standard deviation along the X, Y and Z axis of the accelerometer and the gyroscope signal as features, reporting an accuracy higher than 90%.The Root Mean Square (RMS), minimum, maximum, and zero crossing rate for X, Y, and Z axis were extracted from the accelerometer and gyroscope data, and the ANOVA method was applied for the correct recognition of sitting, resting, turning, and walking with a reported accuracy around 100% [30].The driving, walking, running, cycling, resting, and jogging was recognized by ANN with mean, minimum, maximum, standard deviation, difference between maximum and minimum, Parseval's Energy, Parseval's Energy in the frequency range 0 -2.5 Hz, Parseval's Energy in the frequencies greater than 2.5 Hz, RMS, kurtosis, correlation between axis, ratio of the maximum and minimum values in the FFT, skewness, difference between the maximum and minimum values in the FFT, median of troughs, median of peaks, number of troughs, number of peaks, average distance between two consecutive troughs, average distance between two consecutive peaks, indices of the 8 highest peaks after the application of the FFT, and ratio of the average values of peaks and troughs as features from the accelerometer and gyroscope data, reporting an accuracy between 57.53% to 97.58% [31].The Threshold Based Algorithm (TBA) was applied to the values of the acceleration, and the difference between adjacent elements of the heading, extracted from the accelerometer and gyroscope sensors, in order to recognize going downstairs, going upstairs, running, walking, and jumping with a reported accuracy of 83% [32].The median absolute deviation, minimum, maximum, absolute mean, interquartile range, Signal Magnitude Range, skewness, and Kurtosis were extracted from accelerometer and gyroscope signal for the appl...…”

Pattern recognition techniques for the identification of Activities of Daily Living using mobile device accelerometer

“…With the Bayesian Network, the maximum accuracy reported by the authors is 95.62%; with the Naïve Bayes, the maximum accuracy reported is 97.81%; with the KNN, the maximum accuracy reported is 99.27%; and with the rule based learner, the maximum accuracy reported is 93.53%.Another study [35] presents a solution to recognize walking, jogging, cycling, going up stairs, and going down stairs, implementing a decision tree and a probabilistic neural network (PNN) with some features, such as average of the acceleration, standard deviation for each axis, binned distribution for each axis, and average energy for each axis, reporting results with an average accuracy of 98% with the use of accelerometer.The authors of [36] extracted some features, such as mean, standard deviation, and variance of the accelerometer signal, and implemented the KNN, decision tree, rule-based and MLP methods to recognize walking, sitting, standing, going up stairs and going down stairs activities, verifying that MLP has an accuracy up to 80%.In [37], the authors used the mean, standard deviation, correlation, mean absolute value, standard deviation absolute value, and power spectral density of the accelerometer data, in the Naïve Bayes, KNN, Decision Tree, and SVM methods for the recognition of walking, cycling, running, and standing activities, reporting an accuracy higher than 95%.The authors of [38] implement a method based on the peak values of the accelerometer signal, extraction some features, including the number of peaks every 2 seconds, the number of troughs every 2 seconds, the difference between the maximum peak and the minimum trough every 2 seconds, and the sum of all peaks and troughs, in order to recognize walking, jogging, and marching activities. They implemented the J48 decision tree, bagging, decision table, and Naïve Bayes methods, reporting an accuracy of 94% [38].In [39], the authors implemented a decision tree classifier with several features, such as mean, median, maximum, minimum, RMS, standard deviation, median deviation, interquartile range, energy, entropy, skewness, and kurtosis of the accelerometer data, for the recognition of running, walking, standing, sitting, and laying activities with a reported accuracy of 99.5%.In [40], the authors implemented a SVM method with several features, such as RMS, variance, correlation and energy of the accelerometer data, for the recognition of walking, running, cycling, and hopping with a reported average accuracy of 97.69%.The authors of [41] implemented an ANN with mean, standard deviation, and percentiles of the magnitude of the accelerometer data as features, with a reported accuracy of 92% in the recognition of standing, walking, running, going up stairs, going down stairs, and running.In addition, the authors of [42] implemented an ANN with some features, such as mean, maximum, minimum, difference between maximum and minimum, standard deviation, RMS, Parseval's Energy, correlation between axis, kurtosis, skewness, ratio of the maximum and minimum values in the FFT, difference between the maximum and minimum values in the FFT, median of peaks, median of troughs, number of peaks, number of troughs, average distance between two consecutive peaks, average distance between two consecutive troughs, and ratio of the average values of peaks and troughs based on a window of the accelerometer data. The activities recognized by the method are resting, walking, cycling, jogging, running, and driving [42], r...…”

“…They implemented the J48 decision tree, bagging, decision table, and Naïve Bayes methods, reporting an accuracy of 94% [38].In [39], the authors implemented a decision tree classifier with several features, such as mean, median, maximum, minimum, RMS, standard deviation, median deviation, interquartile range, energy, entropy, skewness, and kurtosis of the accelerometer data, for the recognition of running, walking, standing, sitting, and laying activities with a reported accuracy of 99.5%.In [40], the authors implemented a SVM method with several features, such as RMS, variance, correlation and energy of the accelerometer data, for the recognition of walking, running, cycling, and hopping with a reported average accuracy of 97.69%.The authors of [41] implemented an ANN with mean, standard deviation, and percentiles of the magnitude of the accelerometer data as features, with a reported accuracy of 92% in the recognition of standing, walking, running, going up stairs, going down stairs, and running.In addition, the authors of [42] implemented an ANN with some features, such as mean, maximum, minimum, difference between maximum and minimum, standard deviation, RMS, Parseval's Energy, correlation between axis, kurtosis, skewness, ratio of the maximum and minimum values in the FFT, difference between the maximum and minimum values in the FFT, median of peaks, median of troughs, number of peaks, number of troughs, average distance between two consecutive peaks, average distance between two consecutive troughs, and ratio of the average values of peaks and troughs based on a window of the accelerometer data. The activities recognized by the method are resting, walking, cycling, jogging, running, and driving [42], reporting an accuracy between 57.53% to 97.58%.The authors of [43] implemented a SVN method with mean, minimum, maximum, standard deviation, energy, mean absolute deviation, binned distribution, and percentiles of the magnitude of acceleration as features, with a reported overall accuracy of 94.3% in the recognition of running, staying, walking, going up stairs, and going down stairs.In [44], a method that combines the J48 decision tree, MLP, and Likelihood Ratio (LR) models was implemented and it uses the accelerometer data for the extraction of the minimum, maximum, mean, standard deviation, and zero crossing rate for each axis, and the correlation between axis for the application in the model, in order to recognize the going down stairs, jogging, sitting, standing, going up stairs, and walking activities with a reported accuracy of 97%.The SVM method is also implemented with accelerometer and Global Positioning System (GPS) data for the recognition of walking, standing, and running activities, but the accelerometer features used are minimum, maximum, mean, standard deviation, correlation, and median crossing [68], reporting an accuracy of 97.51%.Another study that uses SVM method makes use of accelerometer, gyroscope, and barometer sensors for the identification of walking, going up stairs, going down stairs, standing, going elevator up, and going elevator down, extraction the mean, mean of 1 st half, mean of 2 nd half, difference of means, slope, variance, standard deviation, RMS, and Signal Magnitude Area [69], reporting an accuracy between 87.45% and 99.25%.The SVM method is also used with the accelerometer, the gyroscope, the barometer, and the GPS sensors...…”

PRO-Fit: A personalized fitness assistant framework

“…A lot of research work has been done in this area, by employing machine learning algorithms on past user activity data [2], heart rate data [3], and accelerometer data [4,5,6] to identify the type of activity, or estimate caloric consumption. In [7], the authors use an on-body chest sensor in coordination with a smartphone to collect the data for the activities performed by the individual, whether static or dynamic.…”