Shanshan
Chen, Christopher L. Cunningham
, John
Lach
UVA Center for Wireless
Health
University of Virginia
BSN, 2011
Extracting
Spatio

Temporal Information
from Inertial Body Sensor Networks for
Gait Speed Estimation
1
Bradford
C.
Bennett,
Research Statement
2
Signal processing challenge to obtain accurate spatial
information from inertial BSNs
Gait speed as an example to extract accurate
spatio

temporal
information
Gait speed is the No. 1 predictor in frailty assessment
require
high gait speed accuracy
d
esire
for continuous, longitudinal gait speed monitoring
Prevailing Technology

for Gait Speed Estimation
Nike+®
Pedometer, cadence
3
Fit

Bit®:
Accelerometer
, cadence
Garmin
Forerunner
®301
Wearable wrist
GPS,
v
elocity
Stopwatch
and Tape
Inertial BSN for Gait Speed Estimation
Portable Solution for Gait Related Analysis
Also provide other
spatio

temporal parameters
Challenges
Spatial Information
Integration
Drift
Mounting Uncertainty
Minimizing Invasiveness
How?
Calculating Stride L
ength/Gait Cycle
Distance =
Stride
Length
of
Gait
Cycle
i
𝑖
=
𝑁
.
𝐶𝑐 𝑠
𝑖
=
1
Average Gait Speed = Distance / Travelling time
4
TEMPO 3.1 inertial BSN platform
developed at the University of Virginia
Contributions
5
Refined human gait model by leveraging biomechanics
knowledge
Improve accuracy without increasing signal processing
complexity
Mounting calibration procedure to correct mounting error
Practical in experiments
Improved gait speed estimation accuracy by combining the
two methods
Outline
6
Current Gait Speed Estimation Method
Gait Cycle Extraction and Integration Drift Cancelation
Stride
Length
Computation by Reference
M
odel
Refined Human Gait Model
Mounting
Calibration
Experiment & Results
Gait Cycle & Integration Drift Cancelation
7
Gyroscope signals on the
sagittal plane
Use foot on ground to find
gait cycle boundaries
Numerically easy to pick
up
–
local maximum
Helpful for canceling
integration drift
Shank
angle is near zero and
does not contribute to the
stride length calculation
when foot is on
ground
Assume
l
inear drift
Stride Length Computation
𝐿
𝐿
:
𝐿𝑔
𝐿𝑛𝑔𝑡ℎ
1
= sin(
θ
2
)
×
𝐿
𝐿
2
=
sin(
θ
2
)
×
𝐿
𝐿
Stride Length
=
1
+
2𝐿
+
1𝐿
+
2
8
Reference Model
S. Miyazaki, “Long

Term Unrestrained Measurement of Stride Length
and
Walking Velocity Utilizing a Piezoelectric Gyroscope”
Outline
9
Current Gait Speed Estimation Method
Gait Cycle Extraction & Integration Drift Cancelation
Stride
Length
Computation by Reference
M
odel
Refined Human Gait Model
Mounting
Calibration
Experiments and Results
Inspection of Gait
P
hase
10
11
𝐿
𝐿
:
𝐿𝑔
𝐿𝑛𝑔𝑡ℎ
,
𝐿
ℎ𝑎
: Shank Length
1
= sin(
θ
𝑖
)
×
𝐿
ℎ𝑎
2
=
sin(
θ
𝑎
)
×
𝐿
𝐿
Stride Length =
1
+
2𝐿
+
1𝐿
+
2
Refined Compound Model
12
Reference Model
Outline
13
Current Gait Speed Estimation Method
Gait Cycle Extraction and Integration Drift Cancelation
Stride
Length
Computation by Reference
M
odel
Refined Human Gait Model
Mounting
Calibration
Experiment & Results
Mounting Calibration
14
Nodes could be rotated 20
°
~30
°
from ideal orientation
Attenuate the signal of interest on the sensitive axis
Essence of Mounting Calibration
Mapping inertial frame (
𝐚
) to global frame(
𝐚
𝐁
) :
𝑹
𝑩𝑰
𝑹
𝑩𝑰
=
𝑩
𝑩
𝑩
Finding
𝑩
,
𝑩
,
𝑩

the x, y, z axis (global frame)
represented by
the inertial frame
Accelerometer readings are the orthogonal bases of
the inertial frame
Ideal Mounting
Non

ideal Mounting
Mounting Calibration Methods
15
Standing straight to get vector
Lift leg and hold still to obtain the rotated
Assumption: rotating only on the sagittal plane, i.e. only y

axis
of accelerometer is rotated, z

axis remain perpendicular to
sagittal plane
Cross product to obtain the third vector
Apply calibration
Validation of Mounting Calibration Algorithm
16
Mounting
Position Rotated
Around Y

axis
Measured by
Proposed Algorithm
Measurement
Error of Angle
0
°

0.072
°
0.072
°
15
°
16.286
°
1.286
°
30
°
27.896
°
2.104
°
45
°
43.954
°
1.046
°
60
°
58.078
°
1.922
°
75
°
74.737
°
0.263
°
90
°
90.461
°
0.461
°
0
0.5
1
1.5
2
2.5
0°
15°
30°
45°
60°
75°
90°
Measurement Error of Angle
Measurement Error of Angle
Pendulum Model to simulate
node rotation on shank
Rotate around z

axis with
controlled degree
Determine the rotation by
Mounting Calibration Algorithm
Achieve an average error of ~1
°
Outline
17
Current Gait Speed Estimation Method
Gait Cycle Extraction and Integration Drift Cancelation
Stride
Length
Computation by reference model
Refined Human Gait Model
Mounting
Calibration
Experiment & Results
Treadmill Control of Speed
Is gait on treadmill
different from on
ground?
Gyroscope signals
collected on
treadmill show no
significant
difference from
those collected on
ground
18
Experiments on Treadmill
Two subjects, a taller male subject and a shorter female subject
Two
trials were conducted for each subject, one with well

mounted nodes and
another with poorly

mounted
nodes to validate mounting calibration
Speeds ranging from 1 to 3 MPH with a 0.2 MPH (0.1m/s) increment for 45
seconds
at each speed
19
Subject with poorly mounted
Inertial BSN nodes performing
mounting calibration on treadmill
Results
Before/After Mounting Calibration
21
•
Badly mounted nodes causes underestimation of gait speed
–
attenuation of
signal due to bad mounting
•
Mounting Calibration has correct the significant estimation error
Before Mounting Calibration
After Mounting Calibration
Results of Two
S
ubjects
22
•
Significantly reduced RMSE compared to the reference model
•
Overestimate
at lower speeds and underestimate at higher
speeds
•
Overestimate
taller
subject’s speeds more than the
shorter subject
Gait Model at Different
S
peeds
The thigh angle can be critical for controlling the step length
23
1
2
1
𝐿
2𝐿
Use thigh nodes to increase accuracy if invasiveness is not a
concern
How accurate is accurate enough?
Depends on application requirement
High Speed
Elimination of thigh angle
results in underestimation of
stride length at high speed
Vice versa at low speed
Results of Two Approaches
24
Double Pendulum at Initial Swing
Single Pendulum Model at Toe

off
•
Better than the reference model
•
Still overestimate the gait speed
Single Pendulum at Toe

Off
Future Work
25
Need more subjects, more gait types, and more gait speeds
For certain types of pathological gait, include those with
shuffling, a wide base, and out

of

plane motion
More refined gait models will be developed based on
biomechanical knowledge
Evaluate if a training set of data can be used to calibrate the
algorithm for each individual subject
Conclusion
26
Achieving an RMSE of 0.09m/s accuracy with a resolution of
0.1m/s
Proposed model shows significant improvement in accuracy
compared to the reference model
M
ounting calibration corrected the estimation error
Leveraging biomechanical domain knowledge simplifies
signal processing
Thanks!
Q&A
Comments 0
Log in to post a comment