A Generalised Approach to Deriving Labelled Transition Systems

Report
Dwaipayan Biswas
University of Southampton, U.K.
ESS Open Day
HIGHLIGHTS OF THE RESEARCH
► Classification of four elementary arm movements that constitute
a significant proportion of daily activities.
► 18 healthy subjects repeating 20 trials of each movement with a
tri-axial accelerometer and a rate gyroscope located proximal to
the wrist.
► Determining the appropriate type of sensor and associated data
processing and classification techniques that may allow long term
monitoring using a body-worn system.
2
HIGHLIGHTS OF THE PAPER
► Ten time domain features extracted on individual sensor
streams, their modulus and specific fused signals.
► Three classifiers used – LDA, QDA and SVM.
► Evaluated using a ‘leave-one-subject-out’ cross validation
strategy.
► The four movements are identified with sensitivities of 83-96%,
using 12 features extracted from individual gyroscope data with
the low-complexity LDA learning algorithm.
3
MOTIVATION
► To develop a wireless body area network (WBAN) system that will detect the use of the impaired
limb during prescribed exercises and activities of daily living (ADL) and classify the type of
movements performed using minimal number of sensors placed at optimal positions on the body.
► Bottleneck ?
Significant energy expenditure in the radio front-end for supporting continuous data transmission.
Reduced battery life of the sensors
► Solution ?
Capturing of vital data by the sensor nodes
Feature extraction and relevant data processing carried out on WSN-nodes itself,
Storage of vital data and features
Transmitting clinically relevant parameters to the patient station at pre-set intervals.
Selecting low-complexity data processing algorithms to extend the battery life since computational
complexity is directly proportional to the energy consumption
4
Experimental Protocol
Movement selection
► Four tasks representative of natural arm movements performed during ADL
 Task A: Reach and retrieve object
 Task B: Lift cup to mouth and return to table
 Task C: Swing arm in horizontal plane through 90° and return
 Task D: Rotate wrist through 90° and return
► Tasks involve sufficiently different kinetics to test the sensors
► 20 trials of each task were performed in 4 groups of 5 repetitions
► Each group of trials was separated by approximately 3 minutes (avoids self
learning)
5
Experimental Protocol
Inertial sensors
► Shimmer kinematic module – 9 degress of freedom (Tri-axial accelerometers
and rate gyroscopes)
 Provides a wireless, wearable solution with minimal hindrance
 Provides a means of time stamping individual events
► Magnetometers can be affected by ferromagnetic materials in the home
environment and hence excluded
► Data acquisition rate – 50 Hz sampling rate
► Accelerometer range : +/- 1.5 g
► Gyroscope range : +/- 500 o/sec
► Position: Wrist - likely to produce the largest sensor response.
6
Data Processing
7
Data Processing
Data Sources
► 6 individual sensor streams – (3 x accelerometers and 3 x gyroscopes)
► 2 modulus signals:
M a  AccX 2  AccY 2  AccZ 2
M g  GyroX 2  GyroY 2  GyroZ 2
► 3 fused signals combining specific accelerometer and gyroscope axes
Movement
Signal Combination
A
AccX * GyroY
B and C
AccY * GyroZ
D
Accz * GyroY
8
Data Processing
Acquisition & Pre-processing
► Raw sensor data is low-pass filtered (3rd order Butterworth filter, cutoff frequency of 12 Hz)
Attenuates the high frequency noise components
► Resultant data is high-pass filtered (3rd order Butterworth filter, cutoff frequency of 0.1 Hz)
Attenuates the low frequency artefacts, e.g. drift
9
Data Processing
Feature extraction
► Time domain features chosen from literature review, including:
Standard Deviation
Maximum Peak Amplitude
Root Mean Square
Absolute Difference (xmax – xmin)
Information Entropy
Index of Dispersion
Jerk Metric
Kurtosis
Peak Number
Skewness
10
Data Processing
Feature normalisation
► Logistic
► Linear
normalisation and standardisation techniques were tested
normalisation produced the best results
xi  xmin
yi 
xmax  xmin
Where yi is the normalised value of xi
11
Data Processing
Feature Selection
approach – various feature combination vectors are selected
to test for the minimum classification error probability
► Wrapper
► Suboptimal
► Selects
searching technique – Sequential forward selection (sfs)
the m best ranked features out of n ranked features (m <= n)
► Chosen
as opposed to other common methods like ReliefF algorithm
and clamping technique which are computationally intensive.
► Features
are selected from:
Individual X, Y, Z axes (3×10 features) for accelerometer and gyroscope,
Modulus signals (1×10 features) for accelerometer and gyroscope and
Fused signals (3×10 features)
12
Data Processing
Classification
► Choice of Classifiers
Linear Discriminant Analysis (LDA)
Quadratic Discriminant Analysis (QDA)
Library for Support Vector Machines (LIBSVM)
► Cross validation methodology
Leave-one-subject-out
► Classification sensitivity of each class (accuracy is not applicable for
multiclass classification)
Computed from confusion matrix (actual versus predicted class)
Diagonal elements represent correctly classified classes
The sensitivity of Class i is given by:
Si 
N (i , i )
c
100
 N (i , j )
j 1
13
Results
Signal
A (%)
B (%)
C (%)
D (%)
Features
Signal
A (%)
B (%)
C (%)
D (%)
Features
Acc_mod
58
58
51
73
9
Acc_mod
49
61
54
72
4
Acc_xyz
85
91
84
90
18
Acc_xyz
89
92
78
91
15
Gyro_mod
82
78
39
80
7
Gyro_mod
82
71
36
85
7
Gyro_xyz
96
83
83
88
12
Gyro_xyz
94
91
95
89
12
Fused
81
74
60
75
13
Fused
86
72
54
74
11
Signal
A (%)
B (%)
C (%)
D (%)
Features
Acc_mod
42
53
55
70
5
Acc_xyz
89
87
82
90
8
Gyro_mod
90
74
35
80
5
Gyro_xyz
97
85
90
89
11
Fused
75
71
50
69
9
Summary of sensitivities for each arm
movement using the modulus signal
(mod), individual sensor signals (xyz)
and fused signals applied to the LDA,
QDA and SVM classifiers.
The number of features required in
each case is also shown.
14
Results
Summary of sensitivities for each arm
movement using the modulus signal
(mod), individual sensor signals (xyz)
and fused signals applied to the LDA,
QDA and SVM classifiers.
The number of features required in
each case is also shown.
15
Conclusion
► The sensitivity for each movement using the individual sensor signals
for both the accelerometer and the gyroscope is better than that for the
fused and the modulus signals.
► The difference in the recognitions rates between modulus and
individual signals is partly due to the fact that bipolar information present
in the raw data is retained with individual sensor signals, but lost with
modulus signals.
► Hence, using the individual sensor signals provides the classifier an
opportunity to select from a larger number of features and hence the
recognition rate for the movements is reflected in the higher accuracies
achieved.
► Considering LDA with individual sensor signals, the gyroscope
recognises the four movements with sensitivities in the range of 83-96%
16
Conclusion
► The accelerometer also has a similar detection rate with sensitivities
in the range of 84-91% across all movements.
► However, the gyroscope uses only 12 features as compared to the 18
used by the accelerometer out of a total of 30 (3×10 features) and
hence is the obvious choice with regard to a lower complexity solution.
► The gyroscope results using individual sensor signals with QDA and
SVM are marginally higher than LDA, though the number of features
required to successfully classify all four arm movements is similar for all
three algorithms
► In view of the trade-off between the recognition rate and the
complexity involved, LDA is considered to be computationally less
complex and hence is the best choice classifier.
17
Thank you for your attention
Any Questions ?
18

similar documents