ClimbTheWorld: Real-time stairstep counting to increase physical activity
-
Upload
university-of-geneva -
Category
Education
-
view
127 -
download
1
Transcript of ClimbTheWorld: Real-time stairstep counting to increase physical activity
04 November 2014 MOBIQUITOUS 2014, London
ClimbTheWorld: Real-time stairstep counting to increase physical activity
Fabio Aiolli, Matteo Ciman, Michele Donini, Ombretta GaggiDepartment of Mathematics,
University of Padua, Italy{aiolli, mciman, mdonini, gaggi}@math.unipd.it
MOBIQUITOUS 2014, London
Health problem
04 November 2014 2 of 25
MOBIQUITOUS 2014, London
Wrong lifestyle
04 November 2014 3 of 25
Sedentary life and bad nutrition are increasing overweight people
Increasing number of diseases like diabetes, cancer etc.
Higher medical costs
MOBIQUITOUS 2014, London
The Fun Theory
04 November 2014 4 of 25
Piano Stairs experiment, Stockholm
MOBIQUITOUS 2014, London
Related works Use mobile and ubiquitous devices to tackle
portability issues Many serious games, gamification systems and
activity recognition to incentivize people to live more actively
Three main problems Fixed position of the smartphone High energy consumption Not always in real-time
04 November 2014 5 of 25
MOBIQUITOUS 2014, London
ClimbTheWorld
04 November 2014 6 of 25
MOBIQUITOUS 2014, London
Goals
Main features: Identify stairstep and distinguish them from
walking step
04 November 2014 7 of 25
A stairstep A step
MOBIQUITOUS 2014, London
Goals
04 November 2014 8 of 25
Main features: Identify stairstep and distinguish them from
walking step Support for (partial) orientation independence Segmentation vs Sliding windows Energy consumption analysis
MOBIQUITOUS 2014, London
Pipeline
04 November 2014 9 of 25
MOBIQUITOUS 2014, London
Data standardization
04 November 2014 10 of 25
DATA STANDARDIZATION
MOBIQUITOUS 2014, London
Orientation independence
Problem: accelerometer data changes depending on the orientation of the smartphone
04 November 2014 11 of 25
MOBIQUITOUS 2014, London
Orientation Independence
First proposal: Mizell in 20031. Take a window of data of fixed time length2. Estimate the gravity component g=(gx , gy , gz)
averaging the readings of the window3. Calculate dynamic component as: d=(ax – gx , ay –
gy , az – gz) for every reading a=(ax , ay , az)4. Calculate vertical component p=5. Horizontal component h = d - p
04 November 2014 12 of 25
MOBIQUITOUS 2014, London
Orientation independence
1. We use a buffer of accelerometer data of the last 500ms
2. We estimate the gravity component g = (gx , gy , gz) as mean value of the buffer readings
3. We calculate the real movement d=(ax – gx , ay – gy , az – gz);
4. Using data from the rotation sensor, we rotate d into d’ to a fixed coordinate system
04 November 2014 13 of 25
MOBIQUITOUS 2014, London
Orientation independence Step 1, 2, 3 of gravity estimation and real
movement estimation are natively supported using the Linear sensor.
04 November 2014 14 of 25
Our solution Native solution
MOBIQUITOUS 2014, London
Segmentation
04 November 2014 15 of 25
TIME FILTER
NO_STAIR
MAYBE_STAIR
SEGMENTATION
MOBIQUITOUS 2014, London
Segmentation A stair step has a specific pattern in
the fixed coordinate system Instead of using sliding window, we
segment data Energy reduction Time becomes a feature Easier learning task User variability
04 November 2014 16 of 25
MOBIQUITOUS 2014, London
Segmentation
04 November 2014 17 of 25
MOBIQUITOUS 2014, London
Features & Classification
04 November 2014 18 of 25
VECTORIAL REPRESENTATION
FEATURES STANDARDIZATIONCLASSIFICATION
STAIR or
NO_STAIR
COUNTING
MOBIQUITOUS 2014, London
Features Basic features to reduce
energy consumption FFT coefficients could be
computationally expensive
74 different values, like average, STD, variance, Signal Magnitude Area
For the Mizell approach, features becomes 74x2 = 148
04 November 2014 19 of 25
MOBIQUITOUS 2014, London
Test Data collected from 7 different users with their
own smartphone 8000 windows, 1500 stairsteps We test Mizell method, Linear method and our
solution at three different frequencies: 20Hz, 30Hz and 50Hz
Learning algorithms: Decision Tree, kNN and KOMD
04 November 2014 20 of 25
MOBIQUITOUS 2014, London
Results
04 November 2014 21 of 25
Mizell Linear Our Method
Mizell Linear Our Method
Mizell Linear Our Method
20Hz 30Hz 50Hz
0.65
0.7
0.75
0.8
0.85
0.9
F-score
DT KNN KOMD
MOBIQUITOUS 2014, London
Energy consumption Energy consumption is a big problem and one of
the most important aspect for final users The best approach is the one that combines low
energy and high precision
04 November 2014 22 of 25
Power Monitor to measure consumed energy
MOBIQUITOUS 2014, London
Energy consumption – Data Stand.
04 November 2014 23 of 25
20Hz 30Hz 50Hz7200
7400
7600
7800
8000
8200
8400
8600
8800
9000
Energy Consumption (uAh)
Mizell Linear Our Method
MOBIQUITOUS 2014, London
Sliding window vs Segmentation
04 November 2014 24 of 25
Sliding window Data segmentation11500
12000
12500
13000
13500
14000
14500
15000
Energy consumption (uAh)
About 1hour saved
MOBIQUITOUS 2014, London
Conclusions Real-time stairstep counter to increase physical
activity during everyday life Main features
Partial support for orientation independence Data segmentation for energy consumption reduction Energy efficiency as key aspect of design
Future works: Use history to increase overall precision of the system Support for trousers pocket
04 November 2014 25 of 25
04 November 2014 MOBIQUITOUS 2014, London
ClimbTheWorld: Real-time stairstep counting to increase physical activity
Fabio Aiolli, Matteo Ciman, Michele Donini, Ombretta GaggiDepartment of Mathematics,
University of Padua, Italy{aiolli, mciman, mdonini, gaggi}@math.unipd.it