G Model

GAIPOS-3881; No. of Pages 3

Gait & Posture xxx (2013) xxxxxx

Contents lists available at SciVerse ScienceDirect

Gait & Posture

journal homepage: www.elsevier.com/locate/gaitpost

Short communication

Validity of the Microsoft Kinect for providing lateral trunk lean

feedback during gait retraining
Ross A. Clark a,*, Yong-Hao Pua c, Adam L. Bryant b, Michael A. Hunt d
a

School of Exercise Science, Australian Catholic University, Melbourne, Australia

Department of Physiotherapy, The University of Melbourne, Melbourne, Australia
c
Department of Physiotherapy, Singapore General Hospital, Singapore
d
Department of Physical Therapy, University of British Columbia, Vancouver, Canada
b

A R T I C L E I N F O

A B S T R A C T

Article history:
Received 22 October 2012
Received in revised form 12 February 2013
Accepted 29 March 2013

Gait retraining programs are prescribed to assist in the rehabilitation process of many clinical conditions.
Using lateral trunk lean modication as the model, the aim of this study was to assess the concurrent
validity of kinematic data recorded using a marker-based 3D motion analysis (3DMA) system and a lowcost alternative, the Microsoft KinectTM (Kinect), during a gait retraining session. Twenty healthy adults
were trained to modify their gait to obtain a lateral trunk lean angle of 108. Real-time biofeedback of the
lateral trunk lean angle was provided on a computer screen in front of the subject using data extracted
from the Kinect skeletal tracking algorithm. Marker coordinate data were concurrently recorded using
the 3DMA system, and the similarity and equivalency of the trunk lean angle data from each system were
compared. The lateral trunk lean angle data obtained from the Kinect system without any form of
calibration resulted in errors of a high (>28) magnitude (mean error = 3.2  2.28). Performing global and
individualized calibration signicantly (P < 0.001) improved this error to 1.7  1.58 and 0.8  0.88
respectively. With the addition of a simple calibration the anatomical position coordinates of the Kinect
can be used to create a real-time biofeedback system for gait retraining. Given that this system is low-cost,
portable and does not require any sensors to be attached to the body, it could provide numerous advantages
when compared to laboratory-based gait retraining systems.
2013 Elsevier B.V. All rights reserved.

Keywords:
Gait training
Biofeedback
Video game
Knee adduction moment
Osteoarthritis

1. Introduction
Methods of implementing gait retraining strategies range from
basic verbal instruction and therapist manual handling through to
real-time instrumented feedback using systems such as markerbased 3D motion analysis (3DMA) [1,2]. A 3DMA system is
particularly suitable for precisely quantifying movement and
providing feedback, however the cost, limited availability and
extensive setup time precludes its use on a large scale in the
clinical setting. Low cost options could include inertial monitoring
sensors such as accelerometers and gyroscopes; however these
sensors possess sources of error such as signal drift and noise
which impede their accuracy.The Microsoft Kinect TM (Kinect) is
low-cost, widely available, and with the release of the ofcial
software development kit (SDK), easily programmed. Evidence
exists that it can accurately assess lateral trunk lean angles during
postural control tests [3] and stride dynamics during walking [4],

and therefore it may be a useful tool for prescribing feedbackassisted gait retraining in a clinical or home setting. Hence, the aim
of this study was to compare data recorded using Kinect and 3DMA
systems during a gait retraining session employing real-time
feedback of performance. The retraining strategy used in this study
was prescribed lateral trunk lean, which has recently been
suggested as a potential gait modication in people with knee
osteoarthritis [1,5].
2. Methods
2.1. Participants
Twenty young, injury free individuals (age: 27.1  4.5 years, height:
173.7  10.3 cm, mass: 71.7  11.0 kg, male = 10) volunteered to participate. Subjects
were required to wear tight-tting shorts and an upper body garment that allowed
placement of reective markers on the upper body. This study was approved by the
institutions Human Research Ethics Committee and all subjects provided informed
consent.
2.2. Procedures

* Corresponding author at: School of Exercise Science, Faculty of Health Sciences,

Australian Catholic University, Victoria, 3095, Australia. Tel.: +61 43173 7609;
fax: +61 3 8344 4188.
E-mail address: ross.clark@acu.edu.au (R.A. Clark).

The gait retraining strategy utilized in this study is similar to that performed
previously [1]. The desired trunk lean angle during the gait retraining session was
set at 108, halfway between the medium (88) and large (128) lateral trunk lean
angles described by Hunt et al. [1], and chosen because it leads to a signicant

0966-6362/$ see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.gaitpost.2013.03.029

Please cite this article in press as: Clark RA, et al. Validity of the Microsoft Kinect for providing lateral trunk lean feedback during gait
retraining. Gait Posture (2013), http://dx.doi.org/10.1016/j.gaitpost.2013.03.029

G Model

GAIPOS-3881; No. of Pages 3

R.A. Clark et al. / Gait & Posture xxx (2013) xxxxxx

reduction in the knee adduction moment. Participants were trained how to

emphasize lateral trunk lean during walking in accordance with Hunt et al. [1]. After
the training was complete, the subjects performed a total of six lateral trunk lean
trials while walking along a 2.5 m walkway starting 3.5 m from the Kinect camera.
Real-time biofeedback of lateral trunk lean was displayed on a laptop computer
with a 43.5 cm screen located next to the Kinect and facing towards the subject at
waist height. Trunk lean data were displayed in real-time with the trace deviating to
the left for left lean to mirror the actual frontal plane trunk movements of the
participant, and with dashed vertical lines at the 08 and 108 angle positions
representing the lateral trunk lean target angle for the testing limb. An example of
this procedure is provided in Fig. 1.
2.3. Data analysis
The 3D skeleton position data for the sternum (SDK output: shoulder centre) and
hip centre (SDK output: hip centre) were used to calculate the lateral trunk lean
angle, and expressed relative to the Kinect camera located directly in front of the
subject and calibrated to the global reference frame of the 3DMA system [3]. These
data were acquired and displayed at their native sampling frequency of
approximately 30 Hz.
The data for the 3DMA system were acquired at 120 Hz using VICON Nexus
V1.5.2 with 12 VICON MX cameras (VICON, UK). Lateral trunk lean angle for this
method was dened as the frontal plane projection of the angle between the global
vertical axis and a line connecting the midpoint of the sternum and the midpoint of
the trans-anterior-superior-iliac-spine, using a marker set similar to the protocol of
Mundermann et al. [5].
The outcome measures for this study were the ability to match the 108 angle
target based on the Kinect data and an inter-device comparison of peak lateral trunk
lean angle during each stride. In addition to the raw joint angle data obtained from
the Kinect, the assessment of two simple calibration techniques was performed to
determine if they improved the accuracy of angle estimation if biases were present,
as observed for some Kinect-derived outcome measures in the study of Clark et al.
[3]. The most basic method consisted of a global calibration, which was created by
dividing the average overall Kinect-derived angle for all subjects by the average
overall 3DMA-derived angle for all subjects and using this as a multiplier for the
individual subject data. This technique would not require any individualized
calibration in future tests, as the global multiplier is the same for all subjects. The
second calibration technique consisted of an individualized calibration, which used
the average of the difference between the Kinect and 3DMA derived angles for the
individual subject as the multiplier essentially a double calibration technique [6].
To examine the practical equivalence between the 3DMA and the Kinect
methods we used two, one-sided t-tests [7] with 90% condence intervals [8] and an

equivalence margin of 28, based on keeping the angle between 88 and 128 as per
Hunt et al. [1]. Additionally, the proportion of the Kinect trials with 28 error was
determined, and we compared the three Kinect calibration methods using exact
McNemar tests for paired proportions with signicance at the two-sided 0.05 level.

3. Results
The peak lateral trunk lean recorded during the trials was
9.7  1.38. This target matching error of 0.38 was similar to the results
of Hunt et al. [1] at prescribed trunk lean angles of 88 (0.348) and 128
(0.888).
The inter-device peak lateral trunk lean angle error results are
provided in Table 1. With no, or only global, calibration the Kinect
values were not equivalent (>28 error) to the 3DMA results (PTOST
values: Raw = 0.99, Global = 0.16). Although the global calibration
was not equivalent, it did result in 85% of the trials having 28
error, which was signicantly (PMcNemar < 0.001) less error than
the data recorded with no calibration. When individualized
calibration was performed, the results were observed to be
equivalent to the 3DMA data, with all subjects within the predetermined 28 range. A separate calibration procedure was also
performed on the results for the rst and second steps independently
to determine whether distance from the Kinect had an inuence on
the calibration value. This was found to result in a minimal (<0.28)
difference in the calculated angle.
4. Discussion
The present study found that, with calibration, a lateral trunk
lean gait retraining program using the skeleton tracking data from
a Kinect compares well with data obtained from a 3DMA system.
Although the raw angle from the skeleton tracking algorithm is
prone to error, performing global or individualized calibration
signicantly improves its accuracy. While the individualized
calibration technique would require subject-specic calibration

Fig. 1. An example of the real-time biofeedback setup during training and gait modication trials. (A) Subjects walked towards the Kinect, which was integrated with a laptop
displaying lateral trunk lean angle (solid white line) as well as the desired 108 target line (vertical dotted line) on a black background. This display is similar to Hunt et al. [1].
(B) A subject with the full body marker set, with the trunk markers used to determine lateral trunk lean angles.

Please cite this article in press as: Clark RA, et al. Validity of the Microsoft Kinect for providing lateral trunk lean feedback during gait
retraining. Gait Posture (2013), http://dx.doi.org/10.1016/j.gaitpost.2013.03.029

G Model

GAIPOS-3881; No. of Pages 3

R.A. Clark et al. / Gait & Posture xxx (2013) xxxxxx

Table 1
Peak lateral trunk lean angle error measured using the three different Kinect calibration methods compared to a 3D motion analysis system (n = 20).
Calibration method

Mean (SD)

90%CI*

PTOST**

Subjects with 28 error (%)

PMcNemary

Raw
Global
Individual

3.2 (2.2)
1.7 (1.5)
0.8 (0.8)

2.44.0
1.32.2
0.71.0

0.99
0.16
<0.001

30
85
100

<0.001
<0.001

SD, standard deviation; CI, condence interval.

*
Equivalence was demonstrated if the 90%CI was totally included within the equivalence interval (28 to 28).
**
P-value from a two one-sided test using an equivalence limit difference of 28.
y
P-value from an exact McNemar test for paired proportions (with reference to the Raw Kinect method).

in future trials, as the multiplier is created based on the subjects

data, it could be easily implemented without the need to use
additional sensors or a 3D camera system. For example, a set 108
angle displayed on the feedback television/computer screen could
be overlaid onto the video image of the person about to perform
the trial. The user could then align their torso with this overlaid
reference, with the difference between the Kinect derived angle
and the known angle then used to create the calibration factor. The
accuracy of this calibration could then be veried manually by the
user via a simple 2D image-based angle assessment, or by a
clinician using a goniometer.
This study indicates that the Kinect can provide accurate
feedback information during gait retraining using a potentially
simple calibration procedure; however the gait retraining strategy
examined possesses heavily exaggerated movement of the trunk.
Whether the accuracy of the Kinect system is sufcient for gait
retraining strategies which require more subtle movements
necessitates further research.
Acknowledgements
The authors would like to thank Karine Fortin and Callan Ritchie
for their assistance with data collection.

Conict of interest
There is no conict of interest.
References
[1] Hunt MA, Simic M, Hinman RS, Bennell KL, Wrigley TV. Feasibility of a gait
retraining strategy for reducing knee joint loading: increased trunk lean guided
by real-time biofeedback. Journal of Biomechanics 2011;44:9437.
[2] Williams G, Clark R, Schache A, Fini NA, Moore L, Morris ME, et al. Training
conditions inuence walking kinematics and self-selected walking speed in
patients with neurological impairments. Journal of Neurotrauma 2011;28:
2817.
[3] Clark RA, Pua YH, Fortin K, Ritchie C, Webster KE, Denehy L, et al. Validity of the
Microsoft Kinect for assessment of postural control. Gait and Posture
2012;36:3727.
[4] Stone E, Skubic M. Evaluation of an inexpensive depth camera for in-home gait
assessment. Journal of Ambient Intelligence and Smart Environments
2011;3:34961.
[5] Mundermann A, Asay JL, Mundermann L, Andriacchi TP. Implications of increased medio-lateral trunk sway for ambulatory mechanics. Journal of Biomechanics 2008;41:16570.
[6] Cappello A, Stagni R, Fantozzi S, Leardini A. Soft tissue artifact compensation in
knee kinematics by double anatomical landmark calibration: performance of a
novel method during selected motor tasks. IEEE Transactions on Biomedical
Engineering 2005;52:9928.
[7] Schuirmann D. On hypothesis testing to determine if the mean of a normal
distribution is contained in a known interval. Biometrics 1981;37:617.
[8] Barker L, Rolka H, Rolka D, Brown C. Equivalence testing for binomial random
variables: which test to use? American Statistician 2001;55:27987.

Please cite this article in press as: Clark RA, et al. Validity of the Microsoft Kinect for providing lateral trunk lean feedback during gait
retraining. Gait Posture (2013), http://dx.doi.org/10.1016/j.gaitpost.2013.03.029