human activity recognition from smart-phone sensor data ... · 10 | human activity recognition from...

Human Activity Recognition from

Smart-Phone Sensor Data using a

Multi-Class Ensemble Learning in

Home Monitoring

HEALTH AND BIO-SECURITY

Soumya GHOSE , Jhimli MITRA 1, Mohan KARUNANITHI 1 and Jason DOWLING 1

1. The Australian e-Health and Research Centre, Commonwealth Scientific and Industrial

Research Organization (CSIRO), Australia

2 | Human Activity Recognition from Smart-Phone Sensor Data using a Multi-Class Ensemble Learning in Home Monitoring

Population Demographics - Australia

Source : https://aifs.gov.au/publications/ageing-yet-diverse


Population Demographics - Australia

The number of Australians aged 65 and over is expected to

increase rapidly from 13 per cent of the population to 25 per cent

of the population by 2042 [1].

Assisting the chronically ill and aged population accounted for

over 70% of Australia's 103.6 billion dollar health expenditure

during 2007-2008, that is only likely to increase [2].

This expenditure may be reduced by remote monitoring of aged

or chronically ill patients that in turn may potentially reduce the

frequency of hospitalizations.1. Australian Government, Australia's Demographic Challenges,

http://demographics.treasury.gov.au/content/_download/australias_demographic_challenges/html/adc-04.asp,

accessed on [14th March, 2015]

2. CSIRO, Home Monitoring of Chronic Disease for Aged Care, http://www.csiro.au/

en/Research/DPF/Areas/Digital-health/Improving-access/Home-monitoring, accessed on [14th March 2015].


Remote Monitoring of Patients

Source : http://www.mdtmag.com/articles/2013/05/wireless-enabled-remote-patient-

monitoring-solutions


Remote Monitoring of Patients - Challenges

Source : http://www.mdtmag.com/articles/2013/05/wireless-enabled-remote-patient-

monitoring-solutions

Data Volume Real-time DecisionsReal Time

Predictive Analysis


Remote Activity Monitoring of Patients – Smartphones

Source: http://www.rcrwireless.com/20150302/opinion/reader-forum-sensor-adoption-in-smartphones-tag10


Smartphone Sensor Data

UCI machine learning repository, Human Activity Recognition

Using Smartphone Dataset

archive.ics.uci.edu/ml/datasets/Human+Activity+Recognition+Using

+Smartphones, accessed on [14th March 2015].

The experiments were carried out on a group of 30 volunteers

within an age bracket of 19-48 years.

Each participant performed six activates (sitting, standing,

walking, laying, walking upstairs and walking downstairs)

wearing a Samsung Galaxy S2 smart-phone.

The experiments were further video tagged to aid in data labelling

(Ground truth).

8 |

Training

Data

Feature

Extraction

Supervised

Learning

Prediction

Model

Test

Data

Feature

Extraction

Real Time

Predictions

Training

Data –Smartphone Sensor Data

Feature

Extraction-Temporal

Motion features

Supervised

Learning –Random Forest

Prediction

Model

Activity Recognition – Building Prediction Model

Human Activity Recognition from Smart-Phone Sensor Data using a Multi-Class Ensemble Learning in Home Monitoring


Activity Recognition – Decision Trees and Forest

http://accidentalepicurean.com/2013/01/accidental-funnies-bacon-decision-tree-chart/

Weak Biased Predictor –

Decision Tree

BMI, Cholesterol, Exercise routine(Good Features)

Ensemble of Decision Trees / Random Forest

= Strong Predictor

Bad Feature


Activity Recognition – Features

The data from the accelerometer and gyroscope, measuring 3-

axial linear accelerations and angular velocity respectively at a

constant rate of 50Hz were captured for activity recognition.

Data pre-processing - The sensor data were pre-processed by

applying a Gaussian noise reduction filter and then re-sampled in

fixed-width sliding windows of 2.56 sec with 50% overlap.

Feature Extraction - From each window, a vector of time and

frequency domains features of the accelerometer signal is

extracted. Mean, standard deviation, signal magnitude area,

entropy, signal-pair correlation, and Fourier Transform were

used to extract the frequency components of the signal.


Activity Recognition – Features

Features - Mean, standard deviation, signal magnitude area,

entropy, signal-pair correlation, and Fourier Transform

(frequency)-> 3-axial linear accelerations and angular velocity,

(sampling resolution 2.56 secs)

Test Data

Prediction (Walking, Sitting etc)


Activity Recognition – Validation

The public dataset was randomly partitioned into 70% training and 30% testing

samples. The common evaluation metrics like the true positive rate, false positive rate

and precision were measured and validated with the ROC curve.


Conclusions

An automatic activity recognition system has been proposed

with a random forest classifier.

In future, this work may be extended to develop a risk

prediction system from patient's activity state to alert health

personnel.

Accurate activity prediction system for the elderly with the

proposed model may potentially reduce the risk associated with

remote monitoring.

Additional sensor data from heart rate monitoring and body

temperature may be incorporated as features in the classification

system to design a patient-specific and illness-specific remote

risk assessment system.


Thank You. Questions?

human activity recognition from smart-phone sensor data ... · 10 | human activity recognition from...

Documents