Triaxial Accelerometer Located on the Wrist for Elderly People’s Fall Detection
1. Triaxial Accelerometer Located on the Wrist for
Elderly People’s Fall Detection
Armando Collado María D. R-Moreno David F. Barrero1
Daniel Rodriguez2
1Computer Engineering Departament, Universidad de Alcalá, Spain 2Computer
Science Departament, Universidad de Alcalá, Spain
IDEAL 2016
Yangzhou, China
Oct. 12-14, 2016
2. Summary
1 Introduction
Motivation
Anatomy of a fall in elderly
Problem statement
2 Data acquisition
Syncope falls
Forward falls
Null class
3 Classifiers trainning
Overview
Time window length
Sample rate
Features selection
4 Model evaluation
Classifiers comparison
Robustness analysis
5 Conclusions and future work
3. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Motivation
Anatomy of a fall in elderly
Problem statement
Introduction
Motivation
Fall detection is a revelant issue in elderly care
... consequence of other health issues
... may originate new health problems
Related work with two approaches
Dedicated devices - Expensive
Cell phones - Psychological rejection
Our proposal: Smartwatches
Rich sensing and communications
Programmable
IDEAL 2016, Yangzhou, China 3 / 18
4. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Motivation
Anatomy of a fall in elderly
Problem statement
Introduction
Anatomy of a fall in elderly
Aging induces physiological changes
Trunk moved forward
Legs apart, slow feet motion
Shifted center of gravity ⇒ Lateral falls highly unusual
Types of falls
Syncope fall: Loss of control over muscles
Forward fall: Trip while walking
IDEAL 2016, Yangzhou, China 4 / 18
5. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Motivation
Anatomy of a fall in elderly
Problem statement
Introduction
Problem statement
Objective
Implement an algorithm able to detect falls with a
smartwatch
Fall detection as a classification task: Fall or not-fall
Data from accelerometer: X, Y and Z
Limited computational resources
We need data to train classifiers
Smartwatch gathering acceleration measures
Hardware imposed the sampling period (20ms)
Different procedures for each type of fall
IDEAL 2016, Yangzhou, China 5 / 18
6. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Syncope falls
Forward falls
Null class
Data acquisition
Syncope falls
Syncope falls
Nursing mannequin
Trainned by experts
Expert supervision
42 simulated falls
30 validated falls
IDEAL 2016, Yangzhou, China 6 / 18
7. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Syncope falls
Forward falls
Null class
Data acquisition
Forward falls
Forward falls
Three volunteers
Trainned by experts
47 simulated falls
40 validated falls
IDEAL 2016, Yangzhou, China 7 / 18
8. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Syncope falls
Forward falls
Null class
Data adquisition
Null class
Syncope fall
Time (s)
Acceleration(m/s^2)
−2001020
AccX
−2001020
AccY
−100102030
1 2 3 4 5
AccZ
Forward fall
Time (s)
Acceleration(ms2
)
−1001020
AccX
−10010
AccY
−20010
1 2 3 4 5 6
AccZ
Not-fall data needed
Vertical motion desirable ⇒ Basketball match
Accelerations lower than a threshold were removed
IDEAL 2016, Yangzhou, China 8 / 18
9. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Overview
Time window length
Sample rate
Features selection
Classifiers tranning
Overview (I)
Data preprocessing
1 Data clean-up
2 Features creation
3 Time windows
construction
Three datasets: Syncope,
Forward and not-fall
Time-window approach
IDEAL 2016, Yangzhou, China 9 / 18
10. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Overview
Time window length
Sample rate
Features selection
Classifiers tranning
Overview (II)
Two binary classifiers
Syncope fall / not-fall
Forward fall / not-fall
Classical classifiers considered: C4.5, 1-NN, Naïve Bayes, PART
10-fold crossvalidation
Attribute Label No. of attributes
Acceleration X AccelX[X1, ...,xN ] N
Acceleration Y AccelY[y1, ...,yN ] N
Acceleration Z AccelZ[Z1, ...,zN ] N
Mean X, Y and Z MeanX, MeanY, MeanZ 3
Std. deviation X, Y and Z DevX, DevY, DevZ 3
Two parameters to set: Window length and sample rate
IDEAL 2016, Yangzhou, China 10 / 18
11. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Overview
Time window length
Sample rate
Features selection
Classifiers trainning
Time window length
Accuracy with varying time window length
Window length Type C4.5 1-NN Naïve Bayes PART
10 (200ms) S 83,71 92,38 68,76 85,18
15 (300ms) S 85,38 93.28 68,62 87,66
20 (400ms) S 87,25 92,86 69,65 89,59
25 (500ms) S 86,10 89,80 71,15 90,85
30 (600ms)* S 87,90 85,49 70,05 91.99
35 (700ms) S 91.59 83,59 71.99 90,31
40 (800ms) S 88,95 82,89 70,64 91,94
10 (200ms) F 86,01 92,84 69,34 88,63
15 (300ms) F 85,38 93,97 74,41 90,34
20 (400ms) F 86.73 94,09 79,1 91,03
25 (500ms)* F 85,87 94.39 80,93 90.91
30 (600ms) F 84,66 92,25 84,58 87,54
35 (700ms) F 82,94 90,59 88.41 85,36
IDEAL 2016, Yangzhou, China 11 / 18
12. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Overview
Time window length
Sample rate
Features selection
Classifiers trainning
Sample rate
Maximum sampling rate imposed by hardware: 20ms
Other rates got by subsampling
Accuracy with different sample rates
Sample
period
Type Attrib C4.5 1-NN Reg. Log. Naïve Bayes PART
20ms S 97 95.38 97.80 90.26 85.71 95.82
40ms S 52 92.10 97.32 87.48 84.35 91.65
60ms S 37 88.38 92.03 87.24 83.60 87.70
20ms F 82 95.66 98.43 88.76 86.74 94.09
40ms F 46 91.97 97.21 89.34 84.43 92.62
60ms F 34 88.45 89.50 84.25 83.46 85.04
Best with high sampling rates
IDEAL 2016, Yangzhou, China 12 / 18
13. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Overview
Time window length
Sample rate
Features selection
Classifiers trainning
Features selection
High number of features
Syncope falls: 30 ∗ 3 + 6 = 96 attributes
Forward falls: 25 ∗ 3 + 6 = 81 attributes
Correlation between attributes and class
Forward fall Syncope fall
% Inf. Attrib. % Inf. Attrib. % Inf. Attrib. % Inf. Attrib.
0.515 DevZ 0.173 AccelX1 0.541 DevY 0.233 AccelX11
0.429 DevY 0.163 MeanZ 0.523 MeanZ 0.232 AccelX8
0.399 MeanX 0.159 AccelZ24 0.238 AccelY12 0.229 AccelY11
0.316 DevX 0.154 AccelX2 0.235 AccelX7 0.229 AccelX9
0.209 MeanY 0.144 AccelZ23 0.235 AccelX12 0.227 AccelX10
0.195 AccelX0 0.135 AccelX3 0.233 AccelX13 0.226 AccelX14
IDEAL 2016, Yangzhou, China 13 / 18
14. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Classifiers comparison
Robustness analysis
Model evaluation
Classifiers comparison
Settings
Sampling rate: 20ms; 30 or 25 samples per window
10-fold crossvalidation
Wrapper feature selection with hill climbing search
C4.5 1-NN Log. Reg. Naïve Bayes PART
Accuracy
S 0.98 1 0.92 0.93 0.96
F 0.98 1 0.89 0.91 0.95
Recall
S 0.98 1 0.94 0.90 0.97
F 0.98 1 0.91 0.92 0.97
Attributes
S 12 7 16 9 7
F 9 13 11 12 7
IDEAL 2016, Yangzhou, China 14 / 18
15. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Classifiers comparison
Robustness analysis
Model evaluation
Robustness analysis
Experiment
Forward falls (not enough mannequins)
Classifiers from previous phase
Evaluate with two unseen people’s falls
C4.5 1-NN Log. Reg. Naïve Bayes PART
Accuracy F 0.91 0.66 0.97 0.98 0.98
Recall F 0.90 0.98 0.95 0.96 0.98
Attributes F 9 13 11 12 7
IDEAL 2016, Yangzhou, China 15 / 18
16. Introduction
Data acquisition
Classifiers trainning
Model evaluation
Conclusions
Conclusions and future work
Mild difficulty
Classifiers implemented on a Samsung Gear S
Satisfactory result
High battery consumption
Future work
Multisensor fall detection (sound and image)
Ensamble of classifiers
IDEAL 2016, Yangzhou, China 16 / 18
17. Thanks for your attention!
谢谢
Scripts and datasets can be freely downloaded from
http://atc1.aut.uah.es/˜david/ideal2016
David F. Barrero
david@aut.uah.es