Fitness Learning

---

Table of Contents

• About
• Data
• Logistic Classifier
• Apple Watch App
• Running the App
• Tools
• Authors
• License
• Acknowledgements

About

A light weight apple watch app with real time recognition of workout categories with pretrained machine learning algorithm.

Motivation

While apple watch provide fitness data monitoring and analysis, it relies on users to manually input the workout category to track. It'll be much more user-friendly if apple watch could automatically detect and categorize the type of workout in the real time.

Data

Data Acquisition

Data was collected directly from the Apple Watch and with the assitance of a third party application called PowerSense,which is a free and powerful motion sensor logging tool that can track and record various sensor data on iPhone and Apple Watch. The main reason we use it is that it provides high sampling rate of data collection. To get more details of the workout, we suggest export the data as XML files, then use the XMLParser to extract interested workout types and export it as csv files for later analysis.

Getting workout data can be simple by using PowerSense:

Setup sampling rate (e.g. 50 Hz) --> click start -- > start workout --> click stop --> export files  

Data Parsing

Data of interest is parsed from the raw xml files export from PowerSense, using our own python module XMLParser:

Dependencies

• pandas
• xml
• numpy
• matplotlib

Initialize

from XMLParser import Parser  
par = Parser("/path/to/file.xml", startDate = "format like 2018-12-01")  

List recorded workout types:

# In: 
par.listTypes  
# Out:
array(['HKCategoryTypeIdentifierAppleStandHour',
       'HKQuantityTypeIdentifierActiveEnergyBurned',
       'HKQuantityTypeIdentifierAppleExerciseTime',
       'HKQuantityTypeIdentifierBasalEnergyBurned',
       'HKQuantityTypeIdentifierDistanceWalkingRunning',
       'HKQuantityTypeIdentifierHeartRate',
       'HKQuantityTypeIdentifierHeartRateVariabilitySDNN',
       'HKQuantityTypeIdentifierRestingHeartRate',
       'HKQuantityTypeIdentifierStepCount',
       'HKQuantityTypeIdentifierWalkingHeartRateAverage'], dtype='<U48')

Extract workout summary:

# In: 
par.loadWorkOutSummary()  
# Out:
                 ActivityType           Duration              EndTime  \
0                   Elliptical  5.665340749422709  2018-03-31 09:49:03   
1                       Rowing  5.387559982140859  2018-03-31 09:58:07   
2  TraditionalStrengthTraining  3.668238099416097  2018-03-31 10:04:51   
3                      Walking  5.929301750659943  2018-03-31 10:14:32   

             StartTime  
0  2018-03-31 09:43:23  
1  2018-03-31 09:52:44  
2  2018-03-31 10:01:11  
3  2018-03-31 10:08:36  

Load a specific workout data:

# In:
par.loadTypeData('HeartRate', plot = False, to_csv = False)
# Out:
#pandas dataframe with columns
Index(['StartTime', 'EndTime', 'HeartRate', 'units'], dtype='object')

Feature Selection

Features saved by apple watch include 'HeartRate', 'ActiveEnergyBurned', 'BasalEnergyBurned', 'DistanceWalkingRunning', and accelerometer measurements (X, Y, Z axis). Explortary data analysis (as in jupyterNotebook) shows that data for 'HeartRate', 'ActiveEnergyBurned', 'BasalEnergyBurned', 'DistanceWalkingRunning' are very noisey and not easily distinguisiable, we focus on the accelerometer data along 3 directions as our selected features and model input. The idea is that accelerometer motion of apple watch should contain enough information for most workouts with arm movement.

Logistic Classifier

A Logistic Regression classifier was trained and tested using the accelerometer data collected to predict the type of activity being performed by the user. The classifier was built in MATLAB and is composed of the following files:

• logistic.m: The main script. Process data, trains and tests a logistic regression model, and outputs a coefficients matrix.

• dataprocess.m: A script that compiles accelerometer data from the .csv files found in \data and splits it into an nxm matrix of features, called X, and an nx1 labels matrix, Y. The matrices are also shuffled before being exported into their corresponding .mat file.

• splitData.m: A function to split matrices X and Y into a training set and a test set, based on the passed 'training ratio'.

• oneVsAll.m: Function trains multiple logistic regression classifiers and returns all the classifiers in a matrix all_theta, where the i-th row of all_theta corresponds to the classifier for label i.

• predictOneVsAll.m: Predict the label for a trained one-vs-all classifier. The labels are in the range 1..K, where K = size(all_theta, 1)

• fmincg.m: Function that allows us to find the minimum point in our cost function.

• lrCostFunction.m: Computes the cost and gradient for logistic regression with regularization.

• sigmod.m: Evaluates the sigmoid function at a given point.

Features and Labels

For our model, we used 13 features, which are the raw accelerometer events logged by the Apple Watch:

• attitude_roll [radians]
• attitude_pitch [radians]
• attitude_yaw [radians]
• rotation_rate (x, y, z)[radians/s]
• gravity (x, y, z)[G],
• user_acc (x, y, z)[G]

Our data consisted of 4 types of activity: rowing, elliptical, push ups, and treadmill, so our outcome labels were classified as follows:

• Elliptical=1, Pushups=2, Rowing=3, Treadmill=4

Accuracy

After splitting the input data into a 70% Training and 30% Test data, our classifier achieved an accuracy of 94.12%

Apple Watch App

We built a light weight Apple Watch app with real time recognition of workout categories with pre-trained machine learning algorithm. It allows exercise aficionados to track their workout without having to enter the exercise set manually.

The app uses Core Motion from Apple's Library, to collect data from the Apple Watch accelerometer and gyroscope plus a machine learning model that can automatically categorize and log an exercise from the collected motion data. As of now, the app can identify the following traditional workouts: Elliptical, rowing machine, treadmill and pushups.

Screenshots

The following are screenshots that show every watch face the user will see before and during the workout.

MyFitnessPal loading screen	View before user starts workout	Start and stop workout button options
Pushups workout view	Elliptical workout view	Rowing workout view

Tools

This project made use of the following tools:

• Apple Watch Development

  • Xcode
  • Swift

• Data Parsing

  • Python
  • Jupyter
  • Matplotlib
  • NumPy

• Logistic Classifier

  • MATLAB

Authors

• Carlos Trejo

  • GitHub
  • LinkedIn
  • Home

• Juan Trejo

  • GitHub
  • LinkedIn
  • Home

• Yu Lu

  • GitHub
  • LinkedIn
  • Home

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Acknowledgements

• PowerSense to collect raw accelerometer vents from the Apple Watch
• Eric Hsiao for providing a template that served as the base for our Apple Watch app