You are on page 1of 9

Journal of Biomechanics 49 (2016) 17421750

Contents lists available at ScienceDirect

Journal of Biomechanics
journal homepage: www.elsevier.com/locate/jbiomech
www.JBiomech.com

Validation of the Leap Motion Controller using markered motion


capture technology
Anna H. Smeragliuolo a,b, N. Jeremy Hill c, Luis Disla d, David Putrino a,b,n
a

Department of Telemedicine and Virtual Rehabilitation, Burke Medical Research Institute, White Plains, NY, USA
Department of Rehabilitation Medicine, Weill-Cornell Medical College, New York, NY, USA
c
Motor Recovery Laboratory, Burke Medical Research Institute, White Plains, NY, USA
d
The City College of New York, School of Biomedical Engineering, New York, NY, USA
b

art ic l e i nf o

a b s t r a c t

Article history:
Accepted 4 April 2016

The Leap Motion Controller (LMC) is a low-cost, markerless motion capture device that tracks hand, wrist
and forearm position. Integration of this technology into healthcare applications has begun to occur
rapidly, making validation of the LMC's data output an important research goal. Here, we perform a
detailed evaluation of the kinematic data output from the LMC, and validate this output against goldstandard, markered motion capture technology. We instructed subjects to perform three clinicallyrelevant wrist (exion/extension, radial/ulnar deviation) and forearm (pronation/supination) movements. The movements were simultaneously tracked using both the LMC and a marker-based motion
capture system from Motion Analysis Corporation (MAC). Adjusting for known inconsistencies in the
LMC sampling frequency, we compared simultaneously acquired LMC and MAC data by performing
Pearson's correlation (r) and root mean square error (RMSE). Wrist exion/extension and radial/ulnar
deviation showed good overall agreement (r 0.95; RMSE 11.6, and r 0.92; RMSE12.4, respectively)
with the MAC system. However, when tracking forearm pronation/supination, there were serious
inconsistencies in reported joint angles (r 0.79; RMSE38.4). Hand posture signicantly inuenced the
quality of wrist deviation (P o0.005) and forearm supination/pronation (Po 0.001), but not wrist exion/
extension (P 0.29). We conclude that the LMC is capable of providing data that are clinically meaningful
for wrist exion/extension, and perhaps wrist deviation. It cannot yet return clinically meaningful data
for measuring forearm pronation/supination. Future studies should continue to validate the LMC as
updated versions of their software are developed.
& 2016 Elsevier Ltd. All rights reserved.

Keywords:
Rehabilitation
Motion capture
Leap motion
Telemedicine
Physical therapy
Motor recovery

1. Introduction
In recent years, a number of low-cost, markerless motionsensing systems have become commercially available for gamers
and hobbyists; the Leap Motion Controller (LMC; Leap Motion Inc.,
San Francisco, CA) is one such system. This device has been
designed to quantify hand movements and gestures. The LMC is
portable, user-friendly, and can reliably track static objects within
0.2 mm accuracy (Weichert et al., 2013). Recently, it has been used
successfully in combination with digital games as a tool for telerehabilitation (Khademi et al., 2014; Putrino, 2014). Telerehabilitation is an emerging method of remote clinical care
delivery that has the potential to signicantly decrease impairment and improve quality of life in individuals suffering from
n
Corresponding author at: Department of Telemedicine and Virtual Rehabilitation, Burke Medical Research Institute, 785 Mamaroneck Ave, White Plains 10605,
NY, USA. Tel.: 1 914 368 3183; fax: 1 914 597 2225.
E-mail address: dap3002@med.cornell.edu (D. Putrino).

chronic disorders of motor control (Garrido et al., 2014; Taylor and


Curran, 2015). Devices such as the LMC have already been established as useful tools for effective telerehabilitation, because they
enable development of interactive systems that make therapy
exercises fun and engaging, and allow therapists to remotely
monitor compliance with ease (Putrino, 2014). It is still unclear,
however, whether the LMC can capture accurate upper-limb
kinematic data in a typical home or clinical environment. Until
this question is addressed, the LMC cannot be used to perform inhome or in-clinic assessments of upper-limb function.
The raw sensor capabilities of the LMC have been validated as
reliable. Work with robotic tools has determined that the LMC can
relay static positional data with a standard deviation o0.5 mm
(Guna et al., 2014). Furthermore, the distance between two moving xed-distance points has been reported within 1.2 mm accuracy (Weichert et al., 2013). However, these studies focus exclusively on the LMC's ability to discriminate end-point motion under
highly controlled, standardized circumstances not its ability to
accurately determine kinematic variables. There are currently no

http://dx.doi.org/10.1016/j.jbiomech.2016.04.006
0021-9290/& 2016 Elsevier Ltd. All rights reserved.

Downloaded from ClinicalKey.com at Nanyang Technological University August 26, 2016.


For personal use only. No other uses without permission. Copyright 2016. Elsevier Inc. All rights reserved.

A.H. Smeragliuolo et al. / Journal of Biomechanics 49 (2016) 17421750

studies that evaluate the LMC's ability to accurately track wrist and
forearm motion in human subjects in a realistic clinical or home
environment. If the LMC can provide accurate joint angle data in
practical settings, it would allow clinicians to remotely perform
assessments of upper-limb function. The implications are enormous for improving the delivery of care to individuals suffering
from motor dysfunction of the wrist and forearm: clinicians could
not only track compliance to home exercise programs, but also
measure the effects of home exercises on joint range of motion
with unprecedented accuracy and regularity. The ability to observe
compliance to a home exercise program alongside measures of
functional improvement will allow clinicians to rigorously evaluate the efcacy of home exercises for each patient.
In order to produce kinematic data, the LMC rst acquires
images of the environment, and uses object recognition to identify
upper limbs in the eld of view. The LMC ofcial website provides
a simple explanation of how it captures images (http://bit.ly/
1A2UI7Q). Following image capture, the LMC software uses a
proprietary variation of stereophotogrammetry for joint motion
inference (Selvik, 1989). Stereophotogrammetry has become a
well-adopted approach to markerless motion capture, but few
devices focus exclusively on deriving upper-limb kinematic data
(Cappozzo et al., 2005; Li et al., 2015). The specic details of how
the LMC optimizes basic stereophotogrammetry algorithms to
produce these data are not publicly available.
For decades, marker-based motion capture systems have been
held as the gold-standard in motion capture technology (Ceseracciu et al., 2014). They are used across disciplines to obtain the
most reliable, non-invasive measurements describing human
motion (Cook et al., 2007). Data from these systems, combined
with joint center estimations established by the use of validated
kinematic algorithms, allow for the calculation of joint motion to a
high degree of accuracy (Metcalf et al., 2008; Todorov, 2007;
Zhang et al., 2011). However, the lengthy setup times and specic
technical knowledge needed not only to operate such a system,
but also to acquire and process the data, renders its use as a
rehabilitation tool unfeasible.
Here, we use gold-standard motion capture technology to
quantify the accuracy with which the LMC records joint angles of
the wrist and forearm, under conditions that were designed to be
reproducible in a supervised clinical environment.
2. Methods
Subjects were recruited from the general population. Inclusion criteria stipulated that subjects must be neurologically healthy, with no history of signicant
injury to either upper limb. Informed written consent was obtained for each subject that was recruited into the study. All experimental practices were conducted
with full approval of the Burke Rehabilitation Hospital Committee for Human
Rights in Research.
2.1. Subject information
We recruited 16 subjects into the study, allowing us to examine 32 hands in
total: 16 right and 16 left. Average hand size was 18.3 cm  9.0 cm (Table 1). There
were six female participants, and ten male participants, with ages ranging from 23
to 55 (mean: 31; standard deviation: 10.1). All subjects completed the assigned
protocol without incident, and no subjects were excluded.

1743

2.2. Markered motion capture system


Eight motion capture cameras from MotionAnalysis Corporation (MAC;
MotionAnalysis Corp., Santa Rosa, CA) were strategically placed in our laboratory to
create a capture volume of approximately 2.2 m  2.3 m  2.3 m. We utilized the
Kestrel camera line for data acquisition, which is capable of a 300 hertz (Hz)
acquisition rate, with 2048  1088 (2.2 million) pixel resolution. The MAC system
allows for highly accurate motion capture data to be collected, which can subsequently be converted to joint angle estimations (Zhang et al., 2011). Using standard
calibration practices outlined by the MAC user's manual, we completed system
calibration prior to data acquisition. Calibration was accepted if average 3D residuals were estimated at under 0.8 mm, and data acquisition was only attempted
after an adequate calibration was achieved. Each subject wore a set of 21 retroreective, spherical (5 mm diameter) markers on their upper body. Marker placement was determined according to private consultations with the MotionAnalysis Corporation as well as prior published recommendations (Rab et al., 2002;
Schmidt et al., 1999). Final placement decisions were made in order to facilitate
ease and accuracy of joint center calculations. The markers were applied to each
anatomical landmark as determined by palpation (Fig. 1).

2.3. The Leap Motion Controller


The LMC is a low-cost, patternless infrared and stereo vision motion capture
device that specializes in markerless motion capture of the forearm, wrist and
hand. It contains two cameras and three infrared lights. It is a small, rectangular
device (13 mm  13 mm  76 mm) that weighs 45 g. It performs live-feedback
motion capture of both hands when it is placed underneath the hands of the
user (Fig. 2). The LMC streams data at a variable acquisition rate of up to 120 Hz. It
is dual platform (Macintosh/Windows), connects to a computer via a USB 3.0 connection, and has a full-functioning Software Developer Kit (SDK). Using the SDK
(v2.3.0), we programmed a piece of data acquisition software that allowed us to
stream and save data from the device.

2.4. Data acquisition


We acquired simultaneous recordings from the LMC and the MAC systems,
while subjects performed a series of one-dimensional rotations of the hand around
axes passing through the wrist: radial/ulnar deviation (rotation about the global xaxis, Fig. 4a), exion/extension (rotation about the global z-axis, Fig. 4b), and
pronation/supination (rotation about the global y-axis, Fig. 4c). Each subject was
instructed to sit in a chair with their arm at their side and elbow exed to 90. A
wooden dowel was placed at the height of each subject's elbow to ensure that they
kept their forearm stationary throughout the protocol. The LMC sat atop a heightadjustable platform, positioned approximately 1.5 hand-lengths below the subject's
hand (Fig. 3). We chose a metric that was customized to each subject so that
subjects with larger hands did not come too close to the sensor, and those with
smaller hands were not too far from the sensor at any point during the protocol.
Hand length, breadth, circumference, and palm length were all measured according
to the US Army guidelines (White, 1980). In order to approximate tracking problems that could arise due to xed, pathological hand postures, each subject
repeated the three motions with their hands in three different positions: open
hand (Fig. 4d), loose st (Fig. 4e), and tight st (Fig. 4f). The subjects completed
each movement pair ve times for the recording. To ensure that movement speed
was consistent across subjects, they were instructed to move to the beat of a
metronome set at 60 beats per minute.

2.5. Kinematic analysis


We used a specialized motion capture software called Cortex (MotionAnalysis
Corp., Santa Rosa, CA) to acquire and process the MAC data. The Skeleton Builder
software package within Cortex allows for calculation of joint centers and a segmental skeleton from the positional marker data. The joint angles presented here
were obtained from a simple three-segment model representing the humerus,
forearm, and metacarpals. This skeleton's local axes were assigned in accordance
with the International Society of Biomechanics' (ISB) recommendations for joint
coordinate systems (Wu et al., 2005). For the right wrist joint, these guidelines

Table 1
Average ( 7 standard deviation), maximum and minimum hand size metrics (cm).

Average
Maximum
Minimum

RH length

RH breadth

RH circ.

RH palm length

LH length

LH breadth

LH circ.

LH palm length

18.3
( 7 1.3)
21.3
16.2

9.1
( 7 0.9)
10.8
8

20.7
(7 1.5)
24.1
8.2

10.0
( 7 0.8)
11.9
8.4

18.2
( 7 1.5)
22
15.6

9.0
( 7 0.9)
11.1
7.7

20.4
( 7 1.6)
23.5
17.8

9.9
( 7 1.0)
11.8
8.2

Downloaded from ClinicalKey.com at Nanyang Technological University August 26, 2016.


For personal use only. No other uses without permission. Copyright 2016. Elsevier Inc. All rights reserved.

1744

A.H. Smeragliuolo et al. / Journal of Biomechanics 49 (2016) 17421750

4
3
5

8
9

7
10
11

12

7
10

11

12

Fig. 1. Upper limb anatomical landmarks for marker placement. 1, 2) the acromion processes; 3) center of the sternum; 4) lateral, and 5) medial humeral epicondyles; 6)
ulnar, and 7) radial styloid processes; 8) the dorsal side of the wrist centered between the styloid processes; 9) center of the third metacarpal (back of the hand); 10) the
middle of the rst metacarpal; 11) the distal phalange of the thumb; 12) the distal phalange of the third (middle) nger.

Fig. 2. (Leap Motion Inc.). Typical use: a user controls virtual hands with the Leap
Motion.

dictate that the y-axis lie pointing proximally along the forearm, the x-axis pointing
volarly, and the z-axis pointing ulnarly (Fig. 5a). These axes were assigned
assuming the wrist's neutral position (i.e. the position at which there is 0 of
rotation) is perpendicular to the oor with the palm facing inward, and thumb
pointing upward. The LMC, on the other hand, considers the wrist to be at 0 of
rotation when the the hand is pronated and the palm is facing the device. As such, a
slight adjustment to the ISB recommendations was necessary in order for our
skeleton's joint angle output to match that of the device. This was achieved by a
simple  90 rotation of the wrist coordinate system about the local y-axis (the axis
running along the forearm; Fig. 5b).

2.6. Temporal alignment of LMC data


Temporal alignment of the raw data was complicated by the LMC's variable
sampling rate: although the MAC system was set to match the nominal 60 Hz
acquisition rate of the LMC, the LMC altered its sampling frequency in an unpredictable manner as a function of time. This resulted in data from the LMC being
desynchronized from the MAC data in a similarly unpredictable way. In order to
properly evaluate the accuracy with which the LMC reports joint angles, we had to
correct for timing discrepancies without altering the amplitude of each signal.
Since variable sampling frequency would not inuence the real-world accuracy of
the measurement of the joint angles, and evaluating the temporal precision of the
LMC was not a priority, we chose to apply a dynamic time warping (DTW) function

Fig. 3. Experimental use: a subject ready to acquire data from the right hand. The
LMC rests on a variable-height platform to the side of the subject, and a wooden
indicator rod is raised to the level of each subject's elbow.

to our data. DTW is an established method for detecting and correcting temporal
alignment discrepancies in time series datasets (Fu et al., 2007). It is commonly
used for purposes such as speech recognition and handwriting verication analyses
(Munich and Perona, 1999; Myers et al., 1980).

Downloaded from ClinicalKey.com at Nanyang Technological University August 26, 2016.


For personal use only. No other uses without permission. Copyright 2016. Elsevier Inc. All rights reserved.

A.H. Smeragliuolo et al. / Journal of Biomechanics 49 (2016) 17421750

1745

Fig. 4. Hand movements: a) radial/ulnar deviation, b) exion/extension, and c) supination/pronation. Hand positions: d) open hand, e) loose st, and f) tight st.

y
z

Fig. 5. Neutral positions of the wrist and corresponding axes. a) ISB recommended neutral wrist position, with grey arrows describing our rotational adjustments. b)
Adjusted neutral wrist position for Leap Motion.
2.7. Comparison of LMC and MAC data
In order to determine whether the LMC tracked some movements more
accurately than others, we performed an ANOVA analysis to look for differences
between the correlation coefcients for all combinations of movement directions
and hand postures. Because we tested across nine different variables, we adjusted
all P-values output by the ANOVA analysis using a Bonferroni Correction. These
analyses can be found in Supplementary Table 1 and Supplementary Figs 13.
In order to quantify the ranges where the LMC showed acceptable performance,
we split the full range of motion (as determined by the MAC measurements) of
each movement direction into multiple small sub-ranges. Supination/pronation
and exion/extension angles were split into 31 sub-ranges of 10 each, and radial/
ulnar deviation into 30 sub-ranges of 5 each. We used three measures to quantify
agreement between LMC and MAC systems:

 Pearson's Rho (r)


Pearson's correlations between LMC and MAC data allowed for a rst-pass
assessment of data agreement for each movement task.

 Bias of the estimator (Bias)


Bias was dened as the difference in means between LMC and MAC data. This
determined whether the LMC generated a constant or less predictable bias.

 Root mean square error (RMSE)

RMSE was calculated as a quantication of the total, non-directional error of


the LMC.

3. Results
3.1. Study participants and initial analysis
We collected a total of 288 simultaneous recordings from 16
subjects using the MAC and LMC systems. Details of left (LH) and
right (RH) hand size metrics, including hand breadth, length, circumference (Circ.), and palm length for all subjects are detailed in
Table 1. We tested wrist and forearm motion in three degrees of
freedom: wrist exion/extension, forearm pronation/supination
and wrist radial/ulnar deviation. The population ranges of motion
for all movements tested using the gold-standard MAC system are
listed in Table 2. The average duration of data recordings was
16.7 7 2.0 s, indicating that movement speed was within parameters that would not affect LMC accuracy.
3.2. Effect of movement direction on correlation of LMC and MAC
The highest correlations were achieved when tracking exion/
extension (r 0.95) and radial/ulnar deviation (r 0.92). However,

Downloaded from ClinicalKey.com at Nanyang Technological University August 26, 2016.


For personal use only. No other uses without permission. Copyright 2016. Elsevier Inc. All rights reserved.

1746

A.H. Smeragliuolo et al. / Journal of Biomechanics 49 (2016) 17421750

Table 2
Average ( 7 standard deviation) joint ranges of motion (degrees).

Range

Extension

Flexion

Pronation

Supination

Radial deviation

Ulnar deviation

44.4 ( 710.3)

64.3 (7 8.4)

11.4 ( 7 9.1)

148.9 ( 735.6)

21.1 ( 7 3.5)

21.5 ( 7 3.4)

Effective Range
30

20
+2 SD

Bias (Degrees)

10

+1 SD

Mean
-1 SD

-10

-2 SD

-20

-30
-40

40
35
30

RMSE

25
20
15
10

+110 to +125

+85 to +95

+95 to +110

+75 to +85

+65 to +75

+55 to +65

+45 to +55

+25 to +35

+35 to +45

+15 to +25

-5 to +5

+5 to +15

-15 to -5

-25 to -15

-35 to -25

-55 to -45

-45 to -35

-65 to -55

-85 to -75

-75 to -65

-95 to -85

-110 to -95

-140 to -125

-125 to -110

Pitch Range (Degrees)

Fig. 6. a) Flexion/extension Bias across range. Negative x-values represent exion for both right and left hands; positive x-values represent extension for both hands. Mean
bias 7 2 standard deviations are delineated on the plot to highlight differing accuracy levels. b) Flexion/extension RMSE across range: Negative x-values represent exion for
both right and left hands; positive x-values represent extension for both hands. Data points attributed to the right hand are represented by warm colors, the left hand by cool
colors. Effective ranges are shaded. The hands represented on the negative and positive sides of the x-axis indicate the directions of movement for both the right and left
hands. (For interpretation of the references to color in this gure legend, the reader is referred to the web version of this article.)

Downloaded from ClinicalKey.com at Nanyang Technological University August 26, 2016.


For personal use only. No other uses without permission. Copyright 2016. Elsevier Inc. All rights reserved.

A.H. Smeragliuolo et al. / Journal of Biomechanics 49 (2016) 17421750

1747

Effective Range
40
30

Bias (Degrees)

20

+2 SD

10

+1 SD
Mean

-1 SD

-10
-2 SD

-20
-30

40
30

RMSE

25
20
15
10
5

+50 to +55

+45 to +50

+35 to +40

+40 to +45

+30 to +35

+20 to +25

+25 to +30

+15 to +20

+10 to +15

0 to +5

+5 to +10

-5 to 0

-10 to -5

-15 to -10

-20 to -15

-30 to -25

-25 to -20

-40 to -35

-35 to -30

-45 to -40

-55 to -50

-50 to -45

-60 to -55

Yaw Range (Degrees)

Fig. 7. a) Radial/ulnar deviation Bias magnitude across range: Negative x-values represent radial deviation of the right hand and ulnar deviation of the left hand; positive xvalues represent ulnar deviation of the right hand and radial deviation of the left. Mean Bias 72 standard deviations are delineated on the plot to highlight differing
accuracy levels. b) Radial/ulnar deviation RMSE across range: Negative x-values represent radial deviation of the right hand and ulnar deviation of the left hand; positive xvalues represent ulnar deviation of the right hand and radial deviation of the left. Data points attributed to the right hand are represented by warm colors, the left hand by
cool colors. Effective ranges are shaded. The hands represented on the negative and positive sides of the x-axis indicate the directions of movement for both the right and left
hands. (For interpretation of the references to color in this gure legend, the reader is referred to the web version of this article.)

the correlation coefcients seen during the forearm pronation/


supination condition were signicantly lower than the other two
movement directions (r 0.79, P o0.001; ANOVA).

3.3. Effect of hand posture on correlation between LMC and MAC


We investigated the effect of three hand postures: open hand,
loose st and tight st on LMC output. For wrist deviation,
performing movements with a tight st posture made the LMC
perform signicantly worse than performing movements with an
open hand (Po 0.007; ANOVA). During forearm pronation/
supination, performing movements with a loose st or tight
st resulted in signicantly lower quality data than performing
the movements with an open hand (P o0.002; ANOVA). LMC
performance during wrist exion/extension was not signicantly
affected by any hand posture (P 0.29; ANOVA).

3.4. When does LMC joint angle inference deteriorate?


For movement in all joint directions, LMC performance seemed
to be dependent on joint ROM (Supplementary Figs 1-3). In order to
track performance of the LMC across a full range of motion, we
monitored Bias, Pearson's correlations, and RMSE (Fig. 6: exion/
extension, Fig. 7: Deviation and Fig. 8: pronation/supination) in
each movement direction. For the presentation of Figs. 68, we
plotted joint angles in the native format of the LMC data output: in
degrees of roll, pitch, and yaw, not anatomical angles. This means
that during deviation, for example, movement in the positive xdirection represents right-hand ulnar deviation and left-hand radial
deviation. In order to recommend an effective range of motion with
which to use the device, we set the threshold for acceptable
agreement at 10% of the full range of motion for each movement.
This threshold was set based upon similar accuracy levels that have
been observed in clinical settings for measuring joint angles with
goniometry (Jonsson and Johnson, 2001; Carter et al., 2009). We

Downloaded from ClinicalKey.com at Nanyang Technological University August 26, 2016.


For personal use only. No other uses without permission. Copyright 2016. Elsevier Inc. All rights reserved.

1748

A.H. Smeragliuolo et al. / Journal of Biomechanics 49 (2016) 17421750

80
60

Bias (Degrees)

40
20

+2 SD
+1 SD
Mean
-1 SD
-2 SD

-20
-40
-60

100
90
80

RMSE

70
60
50
40
30
20

+160 to +180

+180 to +200

+140 to +160

+125 to +140

+95 to +110

+110 to +125

+75 to +85

+85 to +95

+65 to +75

+55 to +65

+35 to +45

+45 to +55

+25 to +35

+15 to +25

-5 to +5

+5 to +15

-15 to -5

-25 to -15

-45 to -35

-35 to -25

-55 to -45

-75 to -65

-65 to -55

-85 to -75

-95 to -85

-110 to -95

-125 to -110

-160 to -140

-140 to -125

-200 to -180

-180 to -160

10

Roll Range (Degrees)

Fig. 8. a) Pronation/supination Bias magnitude across range: Negative x-values represent supination of the right hand and pronation of the left hand; positive x-values
represent pronation of the right hand and supination of the left. Mean Bias 7 2 standard deviations are delineated on the plot to highlight differing accuracy levels. b)
Supination/pronation RMSE across range: Negative x-values represent supination of the right hand and pronation of the left hand; positive x-values represent pronation of
the right hand and supination of the left. Data points attributed to the right hand are represented by warm colors, the left hand by cool colors. No effective range was
identied for this movement. The hands represented on the negative and positive sides of the x-axis indicate the directions of movement for both the right and left hands.
(For interpretation of the references to color in this gure legend, the reader is referred to the web version of this article.)

counted only the region where 10% accuracy was reported for both
the right and the left hands across all hand postures. These effective
ranges are represented by the shaded regions on each of the following graphs. Agreement between the LMC and MAC systems is at
its highest (i.e. LMC Bias is lowest) between  25 and 45 for
wrist exion/extension (Fig. 6a, Po0.001; ANOVA). For radial/ulnar
deviation, the best performance was seen between  10 and 5
(Fig. 7a). The Pearson correlations for wrist exion/extension and
ulnar/radial deviation change expectedly across each movement's
range of motion the correlations are maximized when measuring
angles within each effective range. Hand posture had no signicant
effect on these motions' correlations within the reduced sub-ranges
(P0.54, P 0.99; ANOVA). The associated RMSE across the different sub-ranges also increases accordingly as they approach the
extremes of range (Figs. 6b and 7b, Po0.001; ANOVA). This was of
particular interest because of the large overall range of RMSE values
observed in each movement: exion/extension values ranged from
3.4 to 43.5, radial/ulnar deviation values from 1.6 to 29.9, and
supination/pronation from 11.0 to 76.8 (Figs. 68; Supplementary
Tables 2-4). In the forearm pronation/supination direction of

movement, although overall agreement between systems was fair,


there was no signicant effect on the correlation due to varying
range (P0.18; ANOVA). Furthermore, Bias (Fig. 8a) and RMSE
(Fig. 8b) of the pronation/supination measurements at all ranges
were quite large, not only indicating that the LMC was returning
inaccurate data about forearm rotation, but also failing to establish
an effective range. Finally, the LMC was signicantly less accurate in
tracking supination/pronation when the tight st hand posture
was adopted (Po0.001).

4. Discussion
4.1. The LMC discriminates wrist motion with acceptable accuracy
Devices such as the LMC have the potential to revolutionize
care-delivery for individuals in need of motor rehabilitation of the
hand, due to their accessibility and user-friendly nature (Avola et
al., 2014; Vogiatzaki and Krukowski, 2015; Myers et al., 1980;
Putrino, 2014). The LMC has already proven to be a useful tool for

Downloaded from ClinicalKey.com at Nanyang Technological University August 26, 2016.


For personal use only. No other uses without permission. Copyright 2016. Elsevier Inc. All rights reserved.

A.H. Smeragliuolo et al. / Journal of Biomechanics 49 (2016) 17421750

therapy gaming that can reduce impairment in certain populations


(Garrido et al., 2014). In this study, we believe that the LMC
demonstrated sufcient agreement with the MAC to be used to
quantitatively monitor changes in wrist exion/extension and
ulnar/radial deviation in a supervised clinical environment. We
arrived at this conclusion because the Bias metrics for both
movement directions remained stable (low variance) throughout
the majority of the range of motion. Although there was an offset
between the LMC and MAC data when measuring wrist movements, it was also stable, indicating that it could still be sensitive
to change. Additionally, the fact that starting hand posture mostly
did not signicantly affect the accuracy of joint angle reporting
was encouraging. This indicates that the LMC is versatile enough
to deliver accurate wrist joint angles regardless of the starting
position of the hand, except if it is held in a tight st.
Deterioration of data was noted in all movement directions as
the wrist joint angle moved away from the neutral position.
Although this is a concern in healthy populations, it may not
necessarily affect the device's clinical applicability for the many
conditions that limit wrist mobility. Further examination is
necessary to evaluate the LMC's potential as a clinical measurement tool in specic pathological populations.

1749

our subjects. These markers were essential for MAC motion capture, but are designed specically to reect infrared light. Since the
LMC emits infrared light, the reective properties of the markers
potentially affected the stability of LMC readings. This mostly
occurred during the limits of supination (as this was the hand
movement where the markers were most exposed to the infrared
light of the LMC), so this phenomenon may be additionally
responsible for LMC data instability at the extremes of supination.
A nal, but important, limitation of this study was that,
although we tried to capture natural movement in our subjects,
our protocol was still rather constrained subjects were placed in
a standardized seat, and the LMC was positioned at an optimal
distance from their hand. Subjects received clear instructions on
the protocol, direction and speed of movement were controlled,
and data acquisition was repeated if a deviation from the protocol
was observed. As such, our results are not transferrable to a home
environment. Further research is required to evaluate the role of
the LMC in collecting accurate kinematic data from patients in an
unsupervised environment.

4.4. Future research directions


4.2. Guidelines for clinical use
The results of this study indicate that the LMC can sensitively
track changes in active motion of the wrist, and potentially provide reliable information to a treating clinician about patient
progress while regaining wrist range-of-motion. The LMC is not
sufciently accurate, in any movement direction, to guide unsupervised therapy for clinical cases where certain joint motions are
contraindicated (such as tendon replacement surgeries), or where
range of motion is trying to be determined remotely, without
clinician supervision.
From our study, we can conclude that the effective range of the
LMC as dened by a 10% threshold is  25 to 55 for wrist exion/
extension and  10 to 5 for deviation. Within these ranges, the
LMC showed equal accuracy to the 10% reported by specialized
electronic goniometric devices, and manual goniometric procedures by trained individuals (Jonsson and Johnson, 2001; Carter et
al., 2009; Lowe, 2004). This indicates that the LMC shows noninferiority to techniques that are currently employed in clinical
environments to monitor wrist range-of-motion, and can be
implemented at low-cost. If performing wrist deviation, our ndings suggest that, if possible, the patient should be instructed to
perform movements with an open hand, rather than closing the
hand into a st.
Due to the highly unstable nature of forearm pronation/supination data, we could not recommend any accurate ranges of
motion for this movement.
4.3. Research limitations
Despite acceptable recordings of wrist motion, the LMC
exhibited clear problems while tracking pronation and supination
of the forearm. The most common problem during supination/
pronation was encountered when subjects supinated their forearm
until their hand was parallel with the LMC. In this scenario, the
device was frequently unable to distinguish the volar side of a
right hand from the palmar surface of a left hand (180 from
180 of forearm rotation). Instability of recordings during forearm pronation/supination was a consistent problem across all
subjects, and it signicantly affected the accuracy of reported joint
angles in this plane of motion.
Another factor that may have contributed to unreliable LMC
data was the presence of retro-reective markers on the hands of

The purpose of this study was to evaluate the output of the


LMC in its raw, out-of-the-box form. Our study design only
included healthy subjects, and therefore, before specic recommendations are made, further research is needed to reproduce this
work in varying pathological populations. In future studies, we
also intend to investigate our ability to quantitatively correlate
LMC output to condition-specic functional clinical scales using
machine learning approaches (Barachant et al., 2011; Tuzel et al.,
2008).
Although the LMC showed poor accuracy when tracking forearm
rotation, there are workarounds that can be developed to boost the
accuracy of LMC for clinical applications. For example, when
adapting the LMC to a therapy-gaming context for stroke survivors,
our group applied an algorithm that corrected LMC output in
instances where a tracking error was suspected. Our software would
detect instances where the difference between the current reported
value and the previous value reported by the LMC was unfeasibly
large. In these cases, our algorithm would infer a more feasible joint
position by referencing the previous ve joint positions that had
been reported by the system (Putrino et al., in press). Similarly,
Khademi et al. (2014) have proposed a model that applies inverse
kinematics to the LMC data in order to more accurately infer hand
position. These are only two examples of how a simple custom
development for a clinical environment can make LMC output more
reliable; many more have yet to be examined.
Finally, the LMC development team is constantly producing
new versions of their SDK. Some of these developments contain
updates to their computer-vision software that will continue to
increase accuracy of the system. We anticipate that future software updates will afford greater accuracy to the LMC and should
be investigated. Future studies from our group will focus on
evaluating the feasibility of the LMC as a home-use device by
evaluating the accuracy of kinematic data it returns from patients
using it in unsupervised circumstances.

Conict of interest statement


The authors have no conicts of interest to disclose.

Downloaded from ClinicalKey.com at Nanyang Technological University August 26, 2016.


For personal use only. No other uses without permission. Copyright 2016. Elsevier Inc. All rights reserved.

1750

A.H. Smeragliuolo et al. / Journal of Biomechanics 49 (2016) 17421750

Acknowledgments
The authors would like to acknowledge the Winifred Masterson
Burke Medical Research Institute for the startup funding that
supported this research. The authors also thank the hard-working
Putrino Lab interns: Hannah Actor-Engel, Divija Chopra, David
Harary, Jared Levinson, Matt Marano, Lucas Martins and Cory
Semper.

Appendix A. Supplementary material


Supplementary data associated with this article can be found in
the online version at http://dx.doi.org/10.1016/j.jbiomech.2016.04.
006.

References
Avola, D., Petracca, A., Placidi, G., Spezialetti, M., Cinque, L., Levialdi, S., 2014.
Markerless hand gesture interface based on LEAP motion controller. DMS,
260266.
Barachant, A., Bonnet, S., Congedo, M., Jutten, C., 2011. Multiclass braincomputer
interface classication by Riemannian geometry. IEEE Trans. Biomed. Eng. 59
(4), 920928.
Cappozzo, A., Leardini, A., Chiari, L., 2005. Human movement analysis using stereophotogrammetry: Part 1: theoretical background. Gait Posture 21, 186196.
Carter, T.I., Pansy, B., Wolff, A.L., Hillstrom, H.J., Backus, S.I., Lenhoff, M., Wolfe, S.W.,
2009. Accuracy and reliability of three different techniques for manual goniometry for wrist motion: a cadaveric study. J. Hand Surg. 34, 14221428.
Ceseracciu, E., Sawacha, Z., Cobelli, C., 2014. Comparison of markerless and markerbased motion capture technologies through simultaneous data collection during gait: proof of concept. PloS One 9 (3), e87640.
Cook, J.R., Baker, N.A., Cham, R., Hale, E., Redfern, M.S., 2007. Measurements of wrist
and nger postures: a comparison of goniometric and motion capture techniques. J. Appl. Biomech. 23 (1), 7078.
Fu, A., Keogh, E., Lau, L., Ratanamahatana, C., Wong, R., 2007. Scaling and time
warping in time series querying. VLDB J. 17 (4), 899921.
Guna, J., Jakus, G., Poganik, M., Tomai, S., Sodnik, J., 2014. An analysis of the
precision and reliability of the leap motion sensor and its suitability for static
and dynamic tracking. Sensors.
Jonsson, P., Johnson, P.W., 2001. Comparison of measurement accuracy between
two types of wrist goniometer systems. Appl. Ergon. 32, 599607.
Khademi, M., Mousavi Hondori, H., McKenzie A., Dodakian, L., Lopes, C.V., Cramer, S.
C., 2014. Free-hand interaction with leap motion controller for stroke rehabilitation. In: Proceedings of the Extended Abstracts of the 32nd Annual ACM
Conference on Human Factors in Computing Systems, pp. 16631668.

Li, F., Li, E., Shaee, M.J., Wong, A., Zelek, J., 2015. Dense Depth Map Reconstruction
from Sparse Measurements Using a Multilayer Conditional Random Field
Model. In: Proceedings of the 12th Conference on Computer and Robot Vision
(CRV), 8693.
Lowe, B., 2004. Accuracy and validity of observational estimates of wrist and
forearm posture. Ergonomics 47 (5), 527554.
Metcalf, C.D., Notley, S.V., Chappell, P.H., Burridge, J.H., Yule, V.T., 2008. Validation
and application of a computational model for wrist and hand movements using
surface markers. IEEE Trans. Biomed. Eng. 55 (3), 11991210.
Munich, M., Perona, P., 1999. Continuous dynamic time warping for translation
invariant curve alignment with applications to signature verication. In: Proceedings of the Seventh IEEE International Conference on Computer Vision.
Myers, C., Rabiner, L., Rosenberg, A., 1980. Performance tradeoffs in dynamic time
warping algorithms for isolated word recognition. IEEE Trans. Acoust. 28 (6),
623635.
Navarro, J.E. Garrido, Penichet, V.M. Ruiz, Prez, M.D. Lozano, 2014. Movementbased interaction applied to physical rehabilitation therapies. J. Med. Internet
Res. 16 (12), e281.
Putrino, D., 2014. Telerehabilitation and emerging virtual reality approaches to
stroke rehabilitation. Curr. Opin. Neurol. 27 (6), 631636.
Putrino, D., Zanders, H., Rykman, A., Lee, P., Naeem, O., Disla, L., Panna, T., Edwards,
D., 2016. Use of the GesAircraft video game for upper limb rehabilitation in
stroke: a pilot study. Neurorehabilitation and Neural Repair (in press).
Rab, G., Petuskey, K., Bagley, A., 2002. A method for determination of upper
extremity kinematics. Gait Posture 15 (2), 113119.
Schmidt, R., Disselhorst-Klug, C., Silny, J., Rau, G., 1999. A marker-based measurement procedure for unconstrained wrist and elbow motions. J. Biomech. 32 (6),
615621.
Selvik, G., 1989. Roentgen stereophotogrammetry. Acta Orthop. Scand. 60 (Suppl.
232), S1S51.
Taylor, J., Curran, K., 2015. Using leap motion and gamication to facilitate and
encourage rehabilitation for hand injuries:, Handbook of Research Holistic
Perspectives Gamication Clinical Practice, pp. 183192.
Todorov, E., 2007. Probabilistic Inference of multijoint movements, skeletal parameters and marker attachments from diverse motion capture data. IEEE Trans.
Biomed. Eng. 54 (11), 19271939.
Tuzel, O., Porikli, F., Meer, P., 2008. Pedestrian detection via classication on Riemannian manifold. IEEE Trans. Pattern Anal. Mach. Intell. 30 (10), 17131727.
Vogiatzaki, E., Krukowski, A., 2015. Rehabilitation after stroke using immersive user
interfaces in 3D virtual and augmented gaming environments. EAI Endorsed
Trans. Pervasive Health Technol. 1 (1), e7.
Weichert, F., Bachmann, D., Rudak, B., Fisseler, D., 2013. Analysis of the accuracy and
robustness of the leap motion controller. Sensors 13, 63806393.
White, R.M., 1980. Comparative Anthropometry of the Hand. Army Research and
Development Command, U.S. Natick, Massachusetts Technical Report NATICK/
TR-81/010 (AD A101070).
Wu, G., Van Der Helm, F., Veeger, H., 2005. ISB recommendation on denitions of
joint coordinate systems of various joints for the reporting of human joint
motion Part II: shoulder, elbow, wrist and hand. J. Biomech. 38 (5), 981992.
Zhang, Z., Wong, W., Wu, J., 2011. Ubiquitous human upper-limb motion estimation
using wearable sensors. IEEE Trans. Inf. Technol. Biomed. 15 (4), 513522.

Downloaded from ClinicalKey.com at Nanyang Technological University August 26, 2016.


For personal use only. No other uses without permission. Copyright 2016. Elsevier Inc. All rights reserved.