Defining Accuracy in the use of Kinect v2 for
Exercise Monitoring
Kirk Woolford
School of Arts
University of Surrey
k.woolford@surrey.ac.uk
ABSTRACT
This paper focuses presents a comparison of accuracy and
precision between Kinect v2 (low-cost, markerless) and
Qualisys motion capture (expensive, marker-based system)
for use in healthcare applications. The paper discusses
differences between marker-based and markerless tracking
through an examination of accuracy (closeness of a
measured quantity to its actual value) and precision (or the
proximity of repeated measurements) and the difference
between precisely tracking markers, but inferring the body
structure between them vs directly, but imprecisely tracking
body parts.
The paper discusses issues detracting from the accuracy of
marker-based tracking including human-error in marker-
placement and varying methods of post-processing marker
data to infer joint or bone centers. The paper then presents a
comparison of the Kinect for Windows V2 (K4Wv2)
against a clinical system, and suggests situations where the
Kinect system may be useable in a healthcare environment.
Author Keywords
Motion Capture, Movement Analysis, Kinect, Exercise
Monitoring, Remote Healthcare
ACM Classification Keywords
H.5.2 [Information Interfaces and Presentation]: User
Interfaces—Screen design; Prototyping; User-centered
design; Graphical user interfaces; Theory and methods I.3.6
[Computer Graphics]: Methodology and Techniques—
Interaction techniques; I.3.5
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are not
made or distributed for profit or commercial advantage and that copies bear
this notice and the full citation on the first page. Copyrights for components
of this work owned by others than the author(s) must be honored. Abstracting
with credit is permitted. To copy otherwise, or republish, to post on servers or
to redistribute to lists, requires prior specific permission and/or a fee.
Request permissions from Permissions@acm.org.
MOCO’15, August 14–15, 2014, Vancouver, Canada.
Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-3457-0/15/08…$15.00
DOI: http://dx.doi.org/10.1145/2790994.2791002
112
INTRODUCTION
The release of the original Microsoft Kinect sensor in
November 2010, was accompanied by a wave of fitness
games. These games focused on engaging users in an
attempt to get and keep them exercising. As a result, most
games prioritised responsiveness and rapid feedback over
tracking accuracy. As the health-related games and motion
tracking sensors become more common, healthcare
researchers are exploring the use of sensors such as the
Microsoft Kinect v2 for exercise monitoring and remote
healthcare. However, as these sensors are brought into
clinical and healthcare environments, researchers have
questioned to their accuracy against the more common, and
far more expensive optical systems used in clinical
environments.
Simple, low-cost, portable, sensors such as Kinect are
attractive alternatives to conventional motion capture
systems which are inherently expensive to run as they
require a large dedicated space in which to be used, trained
operators, appropriate hardware and software. Marker-
based systems are accepted as being "clinically-accurate",
but their cost and complexity restrict the settings in which
they can be used. Ideally, these systems would be used
frequently for conditions likely to require long term
exercise monitoring including Parkinson’s disease, cerebral
palsy, stroke and multiple sclerosis, but this would increase
healthcare costs, which are already are a concern across the
world and limit the length and level of long-term feedback a
patient may receive during rehabilitation [1].
The current costs of long term exercise monitoring in the
UK are difficult to quantify due to the varieties of exercise
monitoring available, the range of conditions requiring
exercise monitoring and the differences in services
available in different UK regions. Estimates suggest that the
cost of treating stroke alone across Europe totalled €21,895
million in 2004 and rose to €26,641 million in 2010 [2].
Across Europe €36,619 per traumatic brain injury patient is
spent on rehabilitation. While the proportion of such figures
spent on exercise monitoring and movement rehabilitation
is not available, the average cost of a visit to a
physiotherapist is £36 - 47 and an occupational therapist
£74 (as of 2010-11) [3]. Although the exact cost of exercise
monitoring is not tracked, it is known to be an expensive
aspect of healthcare, and low-cost alternatives are needed.
Novel approaches to exercise monitoring, such as use of
motion tracking sensors developed for video games such as
Microsoft's Kinect, are being seriously considered as tools
for exercise monitoring. From a healthcare provider’s point
of view these systems would allow exercise sessions to be
conducted patients’ homes, which would free up space in
clinic, reduce time spent on system set up, and reduce
overall costs [4]. Benefits for the patient would include
reduced travel time and travel costs, convenient access to
equipment and the possibility of real time feedback and
(a) An inaccurate and imprecise system generates random-like
measurements that are essentially useless in practice.
(c) An accurate and precise system generates identical
measurements that are close to data in the real world.
Unfortunately, 100% accurate and precise systems do not exist in
the real world, because there will always be some error in practice.
Figure 1. Accuracy vs Precision
change to segment angle, with accuracy judged as excellent
based on a Pearson’s product-moment correlation
coefficient (Pearson’s r value) of >0.90 in most movements
measured, but noted a deficiency in the ability of the Kinect
to determine limb rotation movements, similar to other
Kinect accuracy studies [8]. In the monitoring of
Parkinson’s disease patient movement, the Kinect was
found to be accurate (Pearson’s r> 0.9) in measuring timing
and gross spatial movements, such as side stepping
113
interaction. However, many clinicians have concerns about
the accuracy of sensors such as the Kinect. The depth
accuracy of the original Kinect for Xbox360 has been
evaluated extensively showing that it is capable of depth
reconstruction in the order of 1-4 cm at the range of 1-4 m
[5], [6]. The Kinect for Xbox360 has been demonstrated to
return data similar to a VICON MX-based system in
standing reach and balance tests [7]. Diego-Mas & Alcaide-
Marzal investigated the original Kinect sensor through the
measurement of anatomical landmark displacement and
(b) An inaccurate but precise measurement system generates
measurements that are close to each other, but have a systematic
error or bias.
(d) An accurate and modestly precise system generates
measurements that are close to each other and are not
systematically biased with respect to the data in the real world.
This is the level of tracking expected from the Kinect systems.
compared to the measurements of a Vicon system.
However, the accuracy of measuring small movements such
as hand clasping was judged to be very poor, with toe and
finger tapping unable to be detected. Despite this, the
authors concluded that the Kinect, with an overall Pearson’s
r value of .0.80, could potentially be a useful tool for
monitoring individual long term movement changes. [9]
A number of further studies demonstrate that the Kinect for into the skeleton of the person being tracked. However, as
Xbox360 is capable of accurate motion tracking within this a particularly invasive technique and frequently alters
limited ranges and forms of movement [8],[9],[10]. The the movement being tracked, most analysis is performed on
number of caveats and constraints listed in the literature motion captured through markers placed on the skin of the
published on the original Kinect for Xbox 360 and Kinect person being tracked. This skin-placement of markers
for Windows sensors make it difficult to quantify the introduces surface motion artefacts from the movement of a
overall accuracy achieved with the Kinect sensor or to patient’s skin in relationship to their skeleton and soft
compare it directly with industry-standard systems such as tissue. As skin looses its elasticity through age or ill health,
Vicon’s MX series [11]. these artefacts increase. The mass of a tracker introduces
additional movement artefacts. A large tracker, or one on an
extension has significant mass and momentum of its own in
DEFINING ACCURACY addition to the movement of the body-part to which it is
Why is there so much disagreement about the accuracy of attached. The adhesive used to attach a marker to skin may
the Microsoft's Kinect systems? Much of this has to do with also loosen, or may become attached or detached to body
the manner in which accuracy is defined. hair and add extraneous movement, shift position or
become separated from the body during a movement
sequence.
Accuracy vs Precision
As explained in Microsoft’s White paper on Skeletal Joint
Smoothing [13], measurement errors and noise are by-
products of almost any system measuring physical
quantities via sensors. The characteristics of these errors are
normally described as the accuracy and precision of the
system. In such cases, accuracy is defined as the degree of
closeness of a measured quantity to its actual value, and
precision is defined as the degree to which repeated
measurements are close to each other. An accurate system
does not have any systematic error in measurements and
therefore does not add a systematic bias. A precise system
results in measurements close to each other when the
measurement is repeated [13],[14]. The accuracy and
precision concepts are illustrated in Figure 1.
The joint positions returned by the Kinect’s Skeleton
Tracking methods are accurate, meaning that there is no
bias in the joint position data relative to the actual positions
in physical space. In other words, if a person is standing
still, then the average of the joint position data over time is
very close to their positions in physical space. However, the
joint position data is not necessarily precise, meaning that
they samples are scattered around the correct positions in
each frame. In practice, the joint positions for the Microsoft
Kinect are accurate to within a centimeter. [12] This is in
contrast to most clinical optical tracking systems which
measure accuracy in millimeters.

Tracking Joints vs Tracking Markers.

However, whereas optical motion capture systems from
companies such as Vicon and Qualisys are capable of
tracking individual markers with sub-millimeter accuracy,
end users are seldom interested in just knowing the position
of the markers. In clinical movement analysis applications,
the markers are used to track the movement of bones or
centers of joints. In a very limited number clinical
applications, markers can be attached to bone pins inserted
Table 1: Marker placement (Based on Wu et al (2005) and
Kinect V2 landmark/joint description)
In addition to skin and marker movement artefacts, the
placement of markers is complex task prone to human error.
Table 1 and Figures 2 and 3 detail the marker placement
used in tests for this paper. Markers for the scapula were

114
placed using a scapula locator developed at the University These motion artefacts and misplaced markers significantly
of Surrey [15] and depicted in Figure 2. The International offset the precision and accuracy of optical motion capture
Society of Biomechanics (ISB) set out guidelines for in the tracking of discrete markers.
marker placement and joint coordinate systems for use in
upper limb studies [16]. Essentially, these recommended
the use of individual markers on prominent anatomical
features to define the limb length and joint locations, and
markers in the form of clusters, (Figure 3) towards the
middle of the segment to track movement. In order to
reduce the movement of the markers relative to the
underlying bone, the ISB recommends that markers be
placed together in the form of clusters close to the middle
of the segment of interest, while avoiding the muscle belly.
Such segment markers are normally based on either rigid or
deformable frames with a minimum of three markers,
located on the segment of interest. The center of large
bones or complex joints can be particularly difficult to find
through observation or palpitation, i.e., applying pressure to
skin and muscles to determine the location of underlying
bones and other structures.

Figure 3. An acromion cluster as recommended by the

International Society of Biomechanics (ISB) has enough mass
of its own to generate movement artefacts

Markerless Capture
The Microsoft Kinect family of sensors use a very different
approach to track a user's movement from marker-based
systems such as Vicon's MX or Qualisys' ProReflex. Rather
than tracking individual markers and attempting to calculate
the structures between them, the Kinect sensors generate
volumetric silhouettes which are parsed and reduced to
simplified skeletons. Kinect systems attempt to divide
objects located in front of them into segments, compares
these segments against trained decision trees, or decision
forest, to label and position individual body parts. This
Figure 2. Marker placement for tracking of right arm as
process is repeated continually with each segment matched
recommended by the International Society of Biomechanics
(ISB) to the highest probable counterpart in a sample database
and calculating the front, top and left view of the data to
This complex positioning and clustering increases the produce a 3D skeleton extraction from the depth data as
accuracy of optical tracking, but also increases the risk of described by Shotton et al [19] and illustrated in Figure 4.
human error. Accuracy and precision of tracking is based
on the repeatability of marker placement between patients
and tracking sessions. However, inaccurate marker
placements occur frequently in clinical use of optical
motion capture. In their 2011 study, Snider, Degenhardt, et
al B.F., discovered incorrect marker placements on cervical
vertebrae C2, C4 and C6 were in 12.5% of attempts [17].
More disturbingly, Shotton, Fitzgibbon, et al's 2011 study
of marker placement on the lumbar spine found Figure 4. The Kinect V2 process of body part and joint
inaccuracies in 31% of attempts, and concluded that identification (From Shotton et al. 2013)
differences in clinician experience and background, plus
participant characteristics such as obesity and bony
deformity were contributing factors to inaccuracies [18].

115
In other words, the Kinect is running a sequence of very Standardising Frame Rates
rapid pose matching routines across a stream of incoming One area of distinct difference between the Kinect for
depth data in order to make the best guess at the pose of the Windows V2 Software Development Kit (K4Wv2 SDK)
person(s) standing in front of the sensor. The poses are and most optical motion capture systems used in clinical
cross-referenced against a database of millions of images environments is the difference in sampling rates. As the
tagged through a combination of computer vision and K4Wv2 is intended as natural motion user interface, the
manual operation. This tagging process has required human
SDK is optimized to reduce its CPU load and only provide
beings to view and label joints on hundreds-of-thousands of
tracking data when a new pose is detected (i.e. when it
images of human bodies. As a result, where the location of
detects a change in position). The K4Wv2 SDK fires GUI
specific body parts are clearly identifiable, such as with (or NUI) events at unspecified intervals. However, all
wrists, elbows, and knees, the Kinect tends to be very clinical capture position data at set frame rates. Fortunately,
accurate. However, on locations such as shoulders and Microsoft's Open Technology team has worked together
spine, the operator tagging the image has made his/her best- with Open Source developers in the OpenFrameworks and
guess at the centre of the body part. When averaged over Cinder communities to develop an alternative to the
time, these samples do not show a particular bias and are standard K4Wv2 SDK called the Kinect Common Bridge
relatively accurate, but any individual sample will exhibit (KCB). The Kinect Common Bridge allows the Kinect
the lack of precision explained in table 1. sensor to be polled in a timed loop rather than waiting for
the Kinect to fire a NUI event. For the sake of these tests,
the Kinect was sampled at a stable 60 hz.
COMPARING THE KINECT FOR WINDOWS V2 AGAINST
A QUALISYS PROREFLEX SYSTEM
In order to verify the accuracy and precision of the Kinect
for Windows V2 system released in 2014, a number of tests
were conducted at the University of Surrey using a pre-
release system and early betas of the K4Wv2 SDK and a
Qualisys ProReflex marker-based system. These tests were
conducted in collaboration with the University of Surrey's
Dept of Biomedical Engineering. Extensive details of these
tests can be found in Lauren White's MSc Thesis titled, An
Evaluation of the Microsoft Kinect for Windows V2 Motion
Capture Sensor available from the University of Surrey
Library.

Test Protocols (from An Evaluation of the Microsoft

Kinect for Windows V2 Motion Capture Sensor)
The Qualisys ProReflex system is a well-researched formal
motion capture system often referred to as the "gold
standard in accuracy" for clinical motion tracking. Wu et
al's marker and coordinate system [16] based on guidelines
from the International Society of Biomechanics and
detailed in Table 1, was used for the placement of the
Qualisys markers. Three arm movements (shoulder flexion,
shoulder extension and touching the top of the head) were
tracked by both a Kinect V2 sensor and the Qualisys system
simultaneously, at slow and fast self-determined speeds.
The tracking of the shoulder, elbow and wrist joints, and the
lengths of the upper arm and forearm were compared on a
frame by frame and a 0.2 second step (12 frame gap) basis.
A 0.2 second comparison produced larger median errors.
Overall the median mean errors of the Kinect V2 in
tracking the upper limb joints were 1 mm in 1 frame Figure 5. Shoulder movement comparison between Kinect V2
comparisons and between 5-13 mm in 0.2 second and Qualisys Pro Reflex
comparisons.

116
 Figure 6. Elbow and wrist movement comparisons between Kinect V2 and Qualisys Pro Reflex
Time required for marker placement
25 markers were placed on the participants' thorax, arm,
hand, and scapula as described in table 2 and figure 1 and
figure 2. An extended set of markers was placed on
participants and a calibration process was run before several
markers were removed for final capture. The process of
position markes for the Qualisys system took approximately
30 mins per participant. The Kinect, in contrast, took less
than 10 seconds to acquire a tracking subject and begin
returning skeleton data.
117
Results
The mean difference (error) of the Kinect K2 compared to
the Qualisys measurement was calculated for each
movement cycle with the absolute peak difference of each
movement cycle noted in Figures 5 and 6.
The Kinect system regularly measured a smaller movement
in the shoulder joint compared to the Qualisys system
(Figure 5). In 1-frame comparisons the maximum mean
error across all four participants was 3mm (standard
deviation (S.D.) 3mm). The 0.2 sec step-size comparison
produced higher errors, most notably in the fast movements
compared to the 1 frame comparison. The maximum mean
error in 0.2 sec comparison was 34mm (S.D. 38mm). The
peak difference values were considerably higher in the 0.2
sec comparison for most participants. The peak differences
from one subject (001) were noticeably higher than for
other subjects in the 0.2 sec tracking.
In the elbow, unlike the shoulder, the Kinect V2 regularly
over measured the distance moved by the joint (Figure 6).
In 1 frame comparisons the highest mean error was 3mm
(S.D. 11mm). The largest mean error in 0.2 sec comparison
between all other participants was -10mm (S.D. 18mm).
The errors detected at the wrist are similar to those from the
the elbow and more often represent an over measurement
by the Kinect V2 compared to the Qualisys system. The
mean error remained near constant to the elbow, with a
single larger mean error of -2mm (range 8mm). Mean errors
were again higher in the 0.2 sec comparison.
CONCLUSION
Although these tests were conducted on a small sample of
only 5 participants, they verify the tendencies reported in
earlier tests of the first generation of Kinect sensors
[8],[9],[10]. For body parts such as shoulders, whose exact
center of mass cannot be easily identified through sight,
Kinect sensors are accurate, but imprecise. However, for
smaller body parts, whose location can be more easily
determined visually, both generations of Kinect are accurate
and moderately precise. These difficulties in locating the
center of larger body parts appears to be due, at least in
part, to the manner in which Microsoft has constructed their
pose library used for skeleton extraction from volumetric
depth images through the use of manual tagging. Where a
joint's position can be determined as the axis between two
body segments, such as an elbow or wrist, both the
accuracy and precision of the Kinect systems are increased
substantially.
A direct comparison of the accuracy of any generation of
Kinect sensor against marker-based optical motion capture
systems is problematic. Marker-based motion capture
systems are capable of great accuracy in tracking individual
markers, but they are most often used to track body parts
which can only be roughly surrounded by markers. They do
not directly track human skeletons, bones or joints. The
combination of surface movement artefacts, and human
error in marker placement can greatly reduce the accuracy
and precision of marker-based systems. Markerless tracking
systems such as the Kinect avoid these errors as well as the
time and need for expert operators to place makers
accurately and precisely. By removing the trained operator
and the need to place markers, the Kinect can be ready for
capture almost immediately while the marker-based capture
systems frequently take 30 mins or more for marker
placement and calibration.
This ease of set-up and use makes the Kinect a very useful
tool for exercise monitoring. However, the end-user must
be aware of its limitations. The Kinect sensor can
118
accurately track smaller, clearly defined, joints such as
wrists and elbows, but has difficulty locating the exact
centre of larger body areas such as shoulders and spine.
The Kinect can monitor the movement of elbows, and
wrists with high levels of accuracy while the subject is
facing the sensor. Systems like the Kinect cannot match the
accuracy and precision with which individual markers can
be tracked, but for specific movements, they can track a
limited number of bones and joints with similar accuracy
with greatly reduced costs and set up time, while also
providing real-time feedback for users.
ACKNOWLEDGMENTS
Lauren White conducted the comparison of Kinect and
Qualisys movement data as part of her MSc in Biomedical
Engineering programmed under supervision from Professor
David Erwin and Kirk Woolford.
REFERENCES
1. Hadjidj, A., Souil, M., Bouabdallah, A., Challal, Y. &
Owen, H. 2013, "Wireless sensor networks for
rehabilitation applications: Challenges and
opportunities", Journal of Network and Computer
Applications, vol. 36, no. 1, pp. 1-15.
2. Gustavsson, A., Svensson, M., Jacobi, F., Allgulander,
C., Alonso, J., Beghi, E., Dodel, R., Ekman, M.,
Faravelli, C., Fratiglioni, L., Gannon, B., Jones, D.H.,
Jennum, P., Jordanova, A., Jönsson, L., Karampampa,
K., Knapp, M., Kobelt, G., Kurth, T., Lieb, R., Linde,
M., Ljungcrantz, C., Maercker, A., Melin, B.,
Moscarelli, M., Musayev, A., Norwood, F., Preisig, M.,
Pugliatti, M., Rehm, J., Salvador-Carulla, L.,
Schlehofer, B., Simon, R., Steinhausen, H., Stovner,
L.J., Vallat, J., den Bergh, P.V., van Os, J., Vos, P., Xu,
W., Wittchen, H., Jönsson, B. & Olesen, J. 2011, "Cost
of disorders of the brain in Europe 2010", European
Neuropsychopharmacology, vol. 21, no. 10, pp. 718-
779.
3. Curtis, L. 2011, Unit Costs of Health and Social Care
2011, Personal Social Services Research Unit, The
University of Kent , Kent, UK.
4. Chang, Y.-., Chen, S.-. & Huang, J.-. 2011, "A Kinect-
based system for physical rehabilitation: A pilot study
for young adults with motor disabilities", Research in
developmental disabilities, vol. 32, no. 6, pp. 2566-
2570.
5. Jamie Shotton, Toby Sharp, Alex Kipman, Andrew
Fitzgibbon, Mark Finocchio, Andrew Blake, Mat Cook,
and Richard Moore, Real-time human pose recognition
in parts from single depth images, in Communications
of the ACM (CACM), ACM, January 2013.
6. Jamie Shotton, Ross Girshick, Andrew Fitzgibbon, Toby
Sharp, Mat Cook, Mark Finocchio, Richard Moore,
Pushmeet Kohli, Antonio Criminisi, Alex Kipman, and
Andrew Blake, Efficient Human Pose Estimation from
Single Depth Images, in Decision Forests for Computer
Vision and Medical Image Analysis, Springer, 2013.
7. Clark, R.A., Pua, Y., Fortin, K., Ritchie, C., Webster,
K.E., Denehy, L. & Bryant, A.L. (2012a), "Validity of
the Microsoft Kinect for assessment of postural control",
Gait & posture, vol. 36, no. 3, pp. 372-377.
8. Diego-Mas, J.A. & Alcaide-Marzal, J. 2014, "Using
Kinect™ sensor in observational methods for assessing
postures at work", Applied Ergonomics, vol. 45, no. 4,
pp. 976-985.
9. Galna, B., Barry, G., Jackson, D., Mhiripiri, D., Olivier,
P. & Rochester, L. 2014, "Accuracy of the Microsoft
Kinect sensor for measuring movement in people with
Parkinson's disease", Gait & posture, vol. 39, no. 4, pp.
1062-1068.
10. Miranda, D.L., Rainbow, M.J., Crisco, J.J. & Fleming,
B.C. 2013, "Kinematic differences between optical
motion capture and biplanar videoradiography during a
jump-cut maneuver", Journal of Biomechanics, vol. 46,
no. 3, pp. 567-573.
11. Bonnechère, B., Jansen, B., Salvia, P., Bouzahouene,
H., Omelina, L., Moiseev, F., Sholukha, V., Cornelis, J.,
Rooze, M. & Van Sint Jan, S. 2014, "Validity and
reliability of the Kinect within functional assessment
activities: Comparison with standard
stereophotogrammetry", Gait & posture, vol. 39, no. 1,
pp. 593-598.
119
12.Mehran Azimi, (2011) Skeletal Joint Smoothing White
Paper Kinect for Windows 1.5, 1.6, 1.7, 1.8
https://msdn.microsoft.com/en-us/library/jj131429.aspx
13.John Robert Taylor, An Introduction to Error Analysis:
The Study of Uncertainties in Physical Measurements.
University Science Books, 1999.
14.Hughes, Ifan, and Thomas Hase, Measurements and
their Uncertainties: A practical guide to modern error
analysis. Oxford University Press, 2010.
15. Shaheen, A.F., Alexander, C.M. & Bull, A.M.J. 2011,
"Effects of attachment position and shoulder orientation
during calibration on the accuracy of the acromial
tracker", Journal of Biomechanics, vol. 44, no. 7, pp.
1410-1413.
16.Wu, G., van der Helm, F.C.T., (DirkJan) Veeger, H.E.J.,
Makhsous, M., Van Roy, P., Anglin, C., Nagels, J.,
Karduna, A.R., McQuade, K., Wang, X., Werner, F.W.
& Buchholz, B. 2005, "ISB recommendation on
definitions of joint coordinate systems of various joints
for the reporting of human joint motion—Part II:
shoulder, elbow, wrist and hand", Journal of
Biomechanics, vol. 38, no. 5, pp. 981-992. (Gadotti,
Magee 2013)
17. Snider, K.T., Snider, E.J., Degenhardt, B.F., Johnson,
J.C. & Kribs, J.W. 2011, "Palpatory Accuracy of
Lumbar Spinous Processes Using Multiple Bony
Landmarks", Journal of manipulative and physiological
therapeutics, vol. 34, no. 5, pp. 306-313.
18. Shotton, J., Fitzgibbon, A., Cook, M., Sharp, T.,
Finocchio, M., Moore, R., Kipman, A. & Blake, A.
2011a, , Real-Time Human Pose Recognition in Parts
from Single Depth Images. Available:
http://research.microsoft.com/pubs/145347/BodyPartRe
cognition.pdf.