NUI Galway – UL Alliance First Annual ENGINEERING AND - ARAN ...

Determining Subjects’ Activities from Motion Sensor Data using Ensembles
Abstract
Presented here is an approach to predict what type of
motion or activity a person is performing using the
Dynamic Time Warping (DTW) method based on data
from accelerometers worn by or interacted with a
person. DTW is a method of comparing similarities
between two sequences of time series data. When a
particular sequence is tested, it is classified as the
sequence type to which it is most similar. Using
ensembles involves taking the results of multiple
sensors or streams of direction from each sensor i.e. X,
Y, Z, and using the majority decision to be the overall
classification. This technique is being applied to three
separate datasets: 1. The MIT Place Labs, 2. uWave
Application, 3. Cricket Umpire Signals. This is done to
illustrate that the method can be used across a number
of different activity types.
1. Introduction
One of the major motivations for research into
activity detection is to develop methods for real time
monitoring of elderly or post operative patients. “The
number of older persons has tripled over the last 50
years; it will more than triple again over the next 50
years” [1]. Other motivations include the development
of more commercial ventures such as computer games
development. E.g. Nintendo Wii.
For each of the three datasets the subject(s) are
performing multiple activities. The Place Labs contains
information of subjects performing Activities of Daily
Living (ADL) with accelerometers on the wrist, hip and
thigh which include dressing, washing, preparing food,
using phone, computer use and bathroom use. The
uWave and Cricket datasets both have only one
accelerometer and involve hand gestures with a
Nintendo Wii remote/smart phone and signals of a
cricket umpire (accelerometer on wrist), respectively.
Figure 1. uWave [2] & Cricket gestures
[http://crickettips.hpage.com/umpiring__74834427.html]
2. Method
For each activity performed by a subject, the
accelerometer returns the rate of acceleration with a
timestamp. The Place Labs data is then preprocessed
using down sampling, interpolation and moving
Conor O’Rourke
Michael G. Madden
c.orourke3@nuigalway.ie
41
averages where necessary, while the uWave and Cricket
data was received processed [3]. Following this, the
samples are separated into training and test data, they
are then tested with the DTW algorithm. Every test case
is compared with every training case and is classified to
be the type to which it is most similar using the k-NN
algorithm.
Figure 2. DTW matches the similarities between two
sequences of data [4]
3. Ensembles
An ensemble classifier uses a committee of base
classifiers that vote on classification to make a final
decision. For each instance an odd number of different
categories of data from the one instance are tested i.e.
multiple sensors or multiple streams of data (X, Y, Z).
In which case a number of predicted classifications are
made, the majority decision is taken to be the overall
classification.
Previous work on the uWave and Cricket datasets
with DTW achieved results of approx. 72% [3] and
80% [3] respectively. However they have not been
tested using ensembles, it is hoped to improve results
with ensembles using similar DTW methods.
Ensembles have been tested on the Place Labs
dataset which achieved results of 84.3% [5]. In an
attempt to improve these results we plan to optimize the
template selection, sampling frequency and warping
windows.
4. References
[1]. Department of Economic and Social Affairs. World
population Ageing:1950-2050. New York : United Nations
Publications, 2001.
[2]. Liu, J., et al. uWave: Accelerometer-based Personalized
Gesture Recognition. Houston : Rice University and Motorola
Labs, 2008.
[3]. Keogh, E. Unpublished data in private communication.
Nov 2010.
[4]. Ratanamahatana, C.A. and Keogh, E. Making Time-series
Classification More Accurate Using Learned Constraints.
Riverside : University of California - Riverside.
[5]. An Ensemble Dynamic Time Warping Classifier with
Application to Activity Recognition. McGlynn, D. and
Madden, M.G. Cambridge : Thirtieth SGAI International
Conference on Innovative Techniques and Applications of
Artificial Intelligence, 2010.