Wearable Accelerometer Optimal Positions for Human Motion Recognition. The 2020 IEEE 2nd Global Conference on Life Sciences and Technologies (LifeTech 2020), March 10-11, 2020
Call Girls in Bangalore Lavya 💋9136956627 Bangalore Call Girls
Wearable Accelerometer Optimal Positions for Human Motion Recognition(LifeTech2020)
1. Wearable Accelerometer Optimal
Positions for Human Motion
Recognition
2020 IEEELifeTech, Kyoto, March. 10
K e i o U n i v e r s i t y
C h e n g s h u o X i a
Y u t a S u g i u r a
2. • Name: Chengshuo Xia (Nick)
PhD. Candidate
• Affiliation: LifeStyle Computing Lab
(PI: Assist. Prof. Sugiura)
Faculty of Science and Technology
Keio University, Japan
• Interested fields: Human-Computer Interaction
Wearable Technique
Energy Harvesting
Presenter Introduction
2
5. • Wearable sensors have been applied widely to the
recognition of human activities of daily living (ADL),
assists the human daily life from several aspects.
• Also have been constantly focused on from both
commercial perspective and research perspective.
5
[1] Yuan, Ye, and Kris Kitani. "3d ego-pose estimation via imitation learning." Proceedings of the European Conference on Computer Vision (ECCV). 2018.
Background
Huawei Band 4 Pro Google Glass 3D Ego-Pose Estimation[1]
6. • For wearing case, significant issue is to persuade
the user to wear it.
• Thus, the system considering the user’s body
conditions and preferences is necessary.
• For example:
6
Background
Disabled person: Long-term monitoring:
• Disabled body
part may not
suitable for
placing
• Wearing the device for a long
time
7. • Important to study the number of wearable
sensors attached and their positions on the
human body.
• We investigated and presented a series of result
for different numbers and positions of wearable
accelerometers for human ADL recognition.
7
Background
9. • Device: Xsens (MVN Awinda)
• Each unit contains
Accelerometer
magnetometer and gyroscope.
• Sensor positions:
• 17 different locations
(Head/Chest/Waist
RL: hand/Forearm/Shoulder
/Upper leg/Lower leg/Foot)
• Participants: 10
5 males and 5 females
9
Experiment Design
Figure 1. Worn sensors on human body (with portion of practical sensors)
10. • Executed activities:
• Static/Dynamic activity: be performed for 90s;
• Transitional activity: 15times
10
Activity
Activity Type Activity
Static Activity
Standing
Lying
Dynamic Activity
Walking
Running
Going Upstairs
Going Downstairs
Transitional Activity
Sitting-to-standing
Standing-to-sitting
Squatting-to-standing
Standing-to-squatting
11. • Data Processing:
• Machine learning---Support Vector Machine (SVM)[2]
• Data Segmentation:
4s as sliding window size, 2s for overlapping [3]
• Feature Extraction:
• Mean value/Variance/Standard Variance/ 75th percentile/
Inter-percentile;
• Mean value of power spectrum/ Median value of power
spectrum/Shannon entropy value;
• 8 features from time and frequency domain; calculate from 3
axes of accelerometer data;
• Validation: 10-fold cross validation (3 times and
calculate the average value as the accuracy)
11
Support Vector Machine
[2] S. Rosati, G. Balestra, and M. Knaflitz, "Comparison of different sets of features for human activity recognition by wearable sensors," Sensors, vol. 18, p. 4189, 2018.
[3] G. Wang, Q. Li, L. Wang, W. Wang, M. Wu, and T. Liu, "Impact of sliding window length in indoor human motion modes and pose pattern recognition based on
smartphone sensors," Sensors, vol. 18, p. 1965, 2018.
13. • Object/Goal:
Under the requirement of different sensor amount,
figure out the optimal position’s combination
among 17 placed locations.
13
Investigation Object
Worn sensor
number
N∈17
N-Dimension
space
Optimal
sensor
combination
Maximum
classification accuracy
within N-D space
14. • Approach:
Discrete Particle Swarm Optimization (DPSO)based
algorithm;
• Heuristic swarm intelligence algorithm
• Imitate the behaviour of birds foraging
• N-dimension discrete space optimization
We developed a multistage and multi-swarm discrete
particle swarm optimization (MSMS-DPSO) algorithm;
14
MSMS-DPSO Algorithm
Parameters in DPSO Sensor Position Optimization
N-dimensional particle N sensors
Position of a particle Position of sensor (chest/leg/…)
Fitness value Recognition accuracy of activity
Fitness function Relationship between sensor positions and
recognition accuracy
15. • Implementation:
• 3-sensor optimization(as an example):
15
Algorithm Implemetation
Figure 3. Implementation of MSMS-DPSO (3-sensor as an instance)
17
x
x
17
x
x
17
16
15
…
5
x
x
5
x
x
5
x
x
1
2
3
1
x
x
1
x
x
…
Swarm 1 Swarm 3 Swarm 9
Whole population
17
13
7
17
13
7
17
6
9
…
5
6
3
5
7
3
5
6
2
1
4
6
1
4
8
1
3
6
…
Swarm 1 Swarm 3 Swarm 9
Whole populationGlobal optimal particle in each swarm
17
13
7
5
7
3
1
3
6
Swarm 1/Particle 1 Swarm 3/Particle 3 Swarm 9/Particle 9
Whole population
… …
1
3
5
Swarm number:9
Particle number: 27
Swarm number:9
Particle number: 27
Swarm number:9
Particle number: 9
Intragroup optimization end
• 2 stages:
①Intragroup optimization
②Whole swarm optimization
16. • Processing flow:
• Update equations:
16
Algorithm Processing Flow
Figure 2. Working process of MSMS-DPSO
'
1 1(2)i i i
n n nx x v+ += +
1 1 1'
1
1 1 1
[ ] [ ] < 0.5
(3)
[ ] 1 [ ] > 0.5
i i i
n n ni
n i i i
n n n
x if x x
x
x if x x
+ + +
+
+ + +
−
=
+ −
' '
1 1 1 2 2( ) ( )best best
i i i i i i
n n n nv w v c r P x c r G x+ = ⋅ + − + −
Initial solutions
generated
(9*N)
Indicate the first -
dimension position as
2P-1 (P from 1 to 9)
Generate initial fitness
value of each particle
Update the local and
global optimal value in
each swarm
Velocity and position
update (Eq.1 and 2)
Global optimal value
from each swarm as new
particles
Iteration times =
N+1
Generate new global
optimal value
Velocity and position
update (Eq.1 and 2)
All particle converge into
the same position?
Output
1
2
17. • Repetition avoidance:
• 3-sensor optimization (as an instance):
• Bound limitation
• 1≤Position≤17
17
Key Parts for Iterations
1
1
1
2
2
2
3
3
3
4
4
4
5
5
5
6
6
6
15
15
15
16
16
16
17
17
17
...
Converge direction (v < 0)
Global best position [1,2,3]
Current particle position [6,5,4]
1
1
1
2
2
2
3
3
3
4
4
4
5
5
5
6
6
6
15
15
15
16
16
16
17
17
17
...
Global best position [1,2,3]
Current particle position [3,5,4]
Dimension 1
Dimension 2
Dimension 3
Dimension 1
Dimension 2
Dimension 3
After update for
dimension 1
1
1
1
2
2
2
3
3
3
4
4
4
5
5
5
6
6
6
15
15
15
16
16
16
17
17
17
...
Converge direction (v > 0)
Global best position [3,4,6]
Current particle position [1,2,3]
Dimension 1
Dimension 2
Dimension 3
After update for
dimension 1
1
1
1
2
2
2
3
3
3
4
4
4
5
5
5
6
6
6
15
15
15
16
16
16
17
17
17
...
Global best position [3,4,6]
Current particle position [4,2,3]
Dimension 1
Dimension 2
Dimension 3
Figure 4. Position updating for not repeating
(a) Position process for not repeating while v<0 (b) Position process for not repeating while v>0
19. • Configuring the relevant parameters
• N=2/3/4
• Swarm size:9
• Particles in each swarm: 3
• Stop condition:
Intragroup period: Reach iteration times:
N+1;
Whole swarm period: Converge to one
position;
19
Result
• Apply the MSMS-DPSO to investigate 2-sensor,
3-sensor and 4-sensor position combination
Figure 5. Convergence process of MSMD-DPSO (3-sensor example)
Stage 1 Stage 2
20. • Result of MSDS-PSO
20
Result
Sensor
number Position
Accuracy
(%)
1
Right shoulder 88.83%
Waist 87.73%
Left Shoulder 87.68%
2
Waist +Chest 93.55%
Waist+ Head 92.68%
Waist+ Right
shoulder
92.66%
3
Waist + Chest
+Right upper arm
94.57%
Waist + Chest
+Head
94.54%
Waist + Chest
+Left shoulder
94.29%
4
Waist + Chest +
Head +Right upper
arm
95.12%
Waist + Chest +
Head +Left upper
arm
94.83%
Waist + Chest+
Right upper arm
+Left upper arm
94.71%
Acceptable optimal combinations for 1 to 4 sensors:
21. • For different types of activity:
21
Result
0
10
20
30
40
50
60
70
80
90
100
Static Dynamic Transistional
F1-score(%)
Activity Type
Comparison of optimal sensor combinationwith different
number
Right shoulder
Waist+Chest
Waist+Chest+Right upper arm
Waist+Chest+Head+Right upper arm
Static activity: Stand, lie
Dynamic activity: Walk, run, go upstairs, go
downstairs
Transitional activity: sit-to-stand, stand-to-sit, squat-
to-stand, stand-to-squat
Figure 6. F1-score of optimal 1-, 2-, 3- and 4- sensor combinations
• Upper body has advantages
• Significant improvement on
transitional activity recognition
23. 23
Conclusion
• Upper body part, especially the chest, waist,
shoulder and upper arm can present advantages.
• Basically 2 sensors can satisfy the most
situations;
• More sensors used will produce the significant
improvement on transitional activity;
• Future work:
• More complex motions considered;
• More types of sensor considered;
• Rapid algorithm improvement, for online application;