Unsupervised Temporal Segmentation of Skeletal Motion Data using

Joint Distance Representation

Christian Lins

, Sebastian M. Müller

, Max Pﬁngsthorn

, Marco Eichelberg

Alexander Gerka

and Andreas Hein

Carl von Ossietzky University, Ammerländer Heerstr. 140, Oldenburg, Germany

OFFIS - Institute for Information Technology, Escherweg 2, Oldenburg, Germany

Keywords:

Human Motion Analysis, Temporal Segmentation, Joint Distance Matrices, Musculoskeletal Disorders,

Ergonomics Assessment.

Abstract:

In this paper, we present an online method for the unsupervised segmentation of skeletal motion capture data for

the assessment of unfavorable or harmful postures in the context of musculoskeletal disorders. The long-time

motion capture data is segmented into short motion sequences using joint distances of the captured skeleton. We

use the difference between joint distance matrices to detect variances in motion dynamics in which the motion

is separated into either a dynamic motion or a static posture. Then, the static posture can be evaluated using

well-known posture assessment methods such as the Ovako Working postures Analysing System (OWAS) to

derive risk factors for musculoskeletal disorders. The algorithm works in real-time so that it can be incorporated

in live warning systems for unfavorable or harmful postures. We evaluated the segmentation algorithm by

comparing it with results from state-of-the-art ofﬂine motion segmentation algorithms as gold standard. Results

show that the algorithm approaches the performance of state-of-the-art ofﬂine segmentation algorithms.

1 INTRODUCTION

The analysis and classiﬁcation of human motion have

been an active research topic for a long time in va-

rious disciplines (Wang et al., 2003; Aggarwal and

Cai, 1997). Applications include but are not limited

to media animation, biometrics, ergonomics, sports or

computer sciences. In the last years, motion capture

(MC or MoCap) systems, especially such with inertial

sensors (gyroscope, accelerometer, and magnetome-

ter, combined as IMU, short for Inertial Measurement

Unit), have become available to a greater audience.

Such systems, e.g. Xsens MVN (Roetenberg et al.,

2013) or SIRKA (Lins et al., 2015), make it possible

to capture the human motion continuously for hours

or even days. Additionally, such small and embed-

dable motion capture suits make occupational in-situ

observations possible (see Figure 1).

One application is the risk factor assessment

for work-related musculoskeletal disorders (WMSD)

(Wang et al., 2015). Industry workers regularly per-

form unfavorable, harmful or even dangerous postures

during their work shifts. These postures can – together

with other risk factors – lead to musculoskeletal disor-

ders (MSD) such as back pain. Besides the personal

Figure 1: Welder’s digitalized posture and reference footage.

inconveniences, MSDs are a primary cause for sick

leave and early retirement in physically demanding

occupations (Ellegast, 2013). According to (Punnett

and Wegman, 2004), MSDs are the largest category

of work-related illnesses in many countries. The treat-

ment of MSDs causes high costs in the public health

systems of various countries (for example the total

economic costs of MSDs in Canada for the year 1994

equals 3.4% of the gross domestic product) (Coyte

et al., 1998; Walker et al., 2003).

However, the emergence of MSDs can be delayed

or even prevented, if appropriate and timely preventive

measures are applied (Armstrong et al., 1996). Even if

the causes of MSDs are not always occupational cau-

478

Lins, C., Müller, S., Pﬁngsthorn, M., Eichelberg, M., Gerka, A. and Hein, A.

Unsupervised Temporal Segmentation of Skeletal Motion Data using Joint Distance Representation.

DOI: 10.5220/0006598904780485

In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 5: HEALTHINF, pages 478-485

ISBN: 978-989-758-281-3

ses, heavy physical work such as manual handling and

lifting is often considered a risk factor for the emer-

gence of MSD (Amell and Kumar, 2001; Hoy et al.,

2010; Matsui et al., 1997). Thus, prevention measure-

ments become a necessity, e.g. as part of the corporate

health management in industrial companies with phy-

sically hard-working employees. For the prevention,

accompanying physiotherapists will support affected

employees, i.e. those with risk factors, to improve wor-

king processes. This improvement can already prevent

the emergence of occupational diseases, which usually

occur in the second half of life. At the organizational

level, an analysis of the workplace or work planning

could be used as starting point for new managements

concepts that already consider recovery phases in the

work plan.

It is an ongoing task of corporate health manage-

ment to continuously assess psychological and phy-

sical risk factors of every workplace and every wor-

king individual. Risk factors based on physical acti-

vity such as postures and motions can be derived

by human observers using pen-and-paper assessment

methods such as Ovako Working postures Analysing

System (OWAS) (Karhu et al., 1977), Rapid Entire

Body Assessment (REBA) (Hignett and McAtamney,

2000), Rapid Upper Limb Assessment (RULA) (Cor-

lett, 2003), the European assembly worksheet (EAWS)

(Schaub et al., 2013), and others. Such assessments are

tedious and time-consuming, i.e. costly, tasks which

could be supported or replaced by automatized Motion

Capture (MoCap) based systems.

When workers wear a MoCap system during their

work shifts, the posture data captured by the MoCap

system can be either stored for later analysis or prelimi-

narily evaluated on a wearable device (for example as

described by (Nath et al., 2017)). The analyzing soft-

ware can identify critical postures to give the wearer

an immediate feedback on her or his possibly harmful

postures. Then the employee might be able to actively

take a more ergonomic posture or interrupt the work

for a moment to recover.

An evaluating device has to process the motion

data in real-time, a task that requires appropriate algo-

rithms that handle the continuous stream of motions.

Consequently, we propose an approach using an algo-

rithm suited for segmenting dynamic human motion

data into short motion sequences (segments). Our con-

tribution is an online algorithm that uses a time series

of skeletal motions, which is transformed into a joint

distance matrix representation that makes the skeletal

representation rotation and translation invariant. The

algorithm exploits the joint distance representation to

derive a motion dynamics indication, i.e., whether a

motion represents a static posture or a dynamic mo-

tion. The output of the algorithm is a set of segmented

motion sequences with the associated motion dyna-

mics indication that can be further processed by the

assessment system (see also Figure 4).

In this paper, we ﬁrst take a look at related work

of segmentation algorithms (2). We then explain the

segmentation algorithm in detail and compare the ef-

fectiveness of the new algorithm to the segmentation

algorithms of (Krüger et al., 2016; Vögele et al., 2014)

and (Zhou et al., 2013; Zhou et al., 2008). Finally, we

discuss the results and conclude this paper in the last

section.

2 RELATED WORK

Algorithms for unsupervised segmentation of motion

data can be divided into two classes. First, algorithms

that require a priori knowledge, i.e., the complete data

set (or a large batch) to select meaningful segmen-

tation (or cut) points (ofﬂine algorithms). Secondly,

algorithms that can work without (or very little) a pri-

ori knowledge processing data incrementally (online

algorithms).

Segmenting arbitrary time-series data into smaller

parts is a well-known problem in computer science

(Keogh et al., 2001), but there is some work speci-

ﬁc to the problem of segmenting human motion data.

Zhou et al. – as an example of an often mentioned

ofﬂine method – use Dynamic Time Warping (DTW)

to describe the problem of motion segmentation as a

clustering problem which they address using kernel

k-means (Zhou et al., 2008). The authors conclude

that their method is computationally too expensive

for larger motion data sets and provide an improved

method that incorporates a hierarchical decomposition

of motions at different temporal scales (Zhou et al.,

2013). However, the improved version also has a time

complexity of

O(n

max

with

number of input fra-

mes and

number of iterations. So it is presumably

not usable for real-time segmentation.

Vögele et al. , Krüger et al., and Stollenwerk et

al. order the skeletal input data using a k-d-tree to ﬁnd

neighbors of a speciﬁc posture within a search radius

(Vögele et al., 2014; Krüger et al., 2016; Stollenwerk

et al., 2016). A set of neighbors is then represented as a

sparse self-similarity matrix whose graphical structure

is used to ﬁnd distinct motion segments. The method

requires the distance between different neighbored pos-

tures, which are not all available in an online system.

Kulic et al. (Kulic et al., 2009) – in contrast to the

methods mentioned above – describe an approach for

an unsupervised online segmentation and clustering

algorithm. Here, a Hidden Markov Model (HMM)

Unsupervised Temporal Segmentation of Skeletal Motion Data using Joint Distance Representation

479

Figure 2: An example of a kneeling skeleton posture and

a joint distance matrix visualization as grayscale heatmap

(white pixel 0.0, black 1.0.)

is deﬁned on a set of sliding windows containing the

motion data. A standard Gaussian density estimator is

used as observation probability distribution. A modi-

ﬁed Viterbi algorithm is used to ﬁnd the optimum state

sequence in the HMM, which represents the motion

segment. The growing number of states is circumven-

ted by limiting the number of windows, which weakens

the online capabilities.

Koenig and Matari

c (Koenig and Matari

c, 2006)

use a sliding window to search for local maxima in

feature space variances of demonstrated tasks in the

domain of robot learning from demonstration. Fea-

tures can vary depending on the behavior that is de-

monstrated to the robot. The authors recognize the

importance of online capable algorithms for such tasks

and propose a fast and intuitive method for segmenting

tasks into behaviors. Our approach is similar to this

one but speciﬁcally adapted to the purpose of human

motion segmentation. We use distances of concurrent

joint distance matrices as features. Additionally, our

algorithm does not use a sliding but dynamically gro-

wing (up to a maximum for every sequence) window

approach. Koenig and Matari

c’s algorithm emits a

segmentation point when a speciﬁc variance threshold

is recognized, our method, on the other hand, uses the

ratio between the variances of two sequences within

the window as the threshold. We explain our approach

in detail in the next section.

3 MOTION SEGMENTATION

ALGORITHM

3.1 Deﬁnitions and Preconditions

A motion capture system provides a sequence

skeletal postures

ordered in time, so

M =

, S

, ..., S

)

. Every skeletal posture

is an

-tuple

S = (J

, J

, ..., J

)

where every

denotes a joint posi-

tion of the skeleton in R

3.2 Transformation to Distance

Matrices

The use of joint distance matrices as features for clas-

sifying motion capture data was ﬁrst shown by (Vieira

et al., 2012). The representation of a posture as dis-

tance matrix has one major advantage: the representa-

tion is invariant to rotation or translation of the whole

body or the point of view of the observer. That me-

ans that it does not matter whether the skeleton comes

from an inertial sensor suite with its coordinate origin

at the body center or from an optical camera based

system such as Kinect with its coordinate origin at the

observing camera.

The joint distance matrix is also a self-similarity

matrix but it differs from the ones used in the other

mentioned algorithms (Vögele et al., 2014; Krüger

et al., 2016; Stollenwerk et al., 2016) as it is for the

multiple joint coordinates of one pose and not for all

poses of the motion sequence.

The skeletons provided for processing are in most

cases not normalized, i.e., they are not ﬁtted to a stan-

dard skeleton size to make postures comparable. For

this reason, the algorithm can be provided with a nor-

malization factor to transform the skeleton before furt-

her processing. The normalization factor can be deri-

ved from the length of a rigid skeleton segment, e.g.,

shoulder-elbow or a femur bone. A normalization

makes the distance matrices comparable between dif-

ferent motion recordings and subjects. However, for

segmenting motion data into motion sequences, nor-

malization is not necessary.

The distance matrix

of a posture

can be deﬁ-

ned as

= [||J

− J

||]

k,l

(1)

where

l, k

are the joint indices of the skeleton. In

other words,

is a

n ×n

matrix denoting the absolute

distances (a distance metric, e.g. Euclidean) between

every joint in posture

. As mentioned before, the

distances between joints with a rigid connection do

not change during motion so that these values can be

discarded (e.g. set to 0). Figure 2 shows an example

posture and the corresponding joint distance matrix as

grayscale heatmap image.

3.3 Distance Measure of Joint Distance

Matrices

We deﬁne a distance or similarity function

i j

s(D

, D

)

with

d ∈ R

. The similarity measure of

every distance matrix pair

, D

i+1

)

is calculated for

0 ≤ i ≤ m − 1

. That means that we determine the

amount of change between two distance matrices each

HEALTHINF 2018 - 11th International Conference on Health Informatics

480

0 200 400 600 800 1000

0 1 2 3 4 5 6

Frame number

Distance

Figure 3: Distance Measures of Trial 01 with cut points (gray

vertical lines) of our algorithm (l

min

= 3.5).

representing a posture. A plot of a distance change of

one joint over time is shown in Figure 3.

Self-similarity matrices (such as joint distance ma-

trices) have some characteristics that are useful here.

First, the matrix is symmetric on its diagonal. That is

because of the deﬁnition of the distance function that is

used to ﬁll the matrix. Then, if we assume that a bone

connecting two joints is rigid, the distance between

those two does not change throughout the recording.

It will only vary below the noise threshold. As a re-

sult, only the distances of joints not directly connected

are relevant for distance measure between two data

frames. A indicator function

c(J

, J

)

is derived from

the skeleton deﬁnition returning 0 if the connection is

rigid or l = k and 1 if the connection is not rigid.

With this indicator function

we can deﬁne the

distance function as modiﬁed L

norm:

s(X,Y ) =

∑

k=0

∑

l=k

(|x

− y

| · c(J

, J

)) (2)

with

X,Y

being two distance matrices as deﬁned

by Equation 1. Other distance measures for matrices

could be adapted as well.

3.4 Maximize Variance Ratio

Differences between distance matrices (see

Equation 2

) can be seen as a continuous time

series. One problem with sensor-based time series is

the handling of noise. If the data is noisy, a non-robust

segmentation algorithm will return many extrema on

the noisy data. So the challenge is to ﬁnd an algorithm

that properly ﬁlters noise, handles the peculiarities

of human motion and returns timestamps that can be

used as cut points for motion data.

Our algorithm ﬁnds a frame in a dynamic window

that separates the window into two segments, one

with high variance and one with low variance. Such

frame can be seen as start or end point of a motion

segment, e.g. a constrained posture with little dynamic

and a change to high dynamic (meaning high variance)

when the subject starts to move.

Algorithm 1: Segmentation algorithm.

Require: W

is a

-element window of distances

Require: l

min

, l

max

is the minimum/maximum seg-

ment length (default: l

max

= 4 · l

min

)

Require: T

is the min. threshold ratio that is required

for segmenting a window (default: T = 10)

1: function SEGMENT(W, l

min

, l

max

, T )

2: r ← 0

3: for p ← l

min

to p ← n − l

min

4: µ

← mean(W, 0, p)  Sample mean

5: µ

← mean(W, p + 1, n)

6: σ

← var(W, µ

, 0, p)  Sample variance

7: σ

← var(W, µ

, p + 1, n)

8: r ← max(r,





max(σ

, σ

)

min(σ

, σ

)





)

9: if r > T or length(W ) > l

max

then

10: p ← ﬁndMin(p,W, σ

, σ

, l

min

)

11: return p and σ

12: return  No proper cut point found yet

The algorithm is outlined in pseudocode as Algo-

rithm 1. We assume a minimum segment length

min

avoid segmentation into very small motion fragments.

So the ﬁrst possible segmentation (or cut) point is the

index at l

min

The index

for which the ratio

r = max

is a plau-

sible candidate for a cut point. To ensure that

a local minimum, which can be seen as a rest pose,

we search

steps for a local minimum in the part of

the window with lower variance.

must be a small

fraction of

. The function ﬁndMin (see Algorithm 1)

implements this linear search within the surrounding

data points (e.g. p −

→ p +

3.5 Detection of Constrained Postures

Constrained postures are often performed by industry

workers during their duties, e.g. holding a tool in a

static awkward position for several minutes. Such pos-

tures should be avoided as they can cause damage to

the musculoskeletal system. As a side effect of the

segmentation algorithm, the variance of the motion

Unsupervised Temporal Segmentation of Skeletal Motion Data using Joint Distance Representation

481

Motion Segmentation

Skeletal

Motion

Capture

Data

Joint Distance

Representation

Temporal

Motion

Segmentation

Algorithm

Posture Assessment

OWAS

EAWS, ...

Posture Risk

Evaluation

Risk

Report

Static

Postures

Motion

Segments

Figure 4: Process of WMSD Risk Factor Identiﬁcation as enhancement to the model of (Lins et al., 2016).

segment

is known and can be used to detect sta-

tic postures, because the high variance in the joint

distances is caused by rapid movement of limbs. A

small variance in the joint distances of consecutive

frames can be caused by no or little overall movement

of the subject or by continuous movement with very

little change over time (e.g. sneaking). As a result,

a segment can be identiﬁed as static posture by the

segmentation algorithm, but its classiﬁcation as con-

strained posture must be done by other parts of the

assessment system (see Posture Assessment in Figure

4).

3.6 Time Complexity

For every input frame that is added to the current win-

dow

the segment function (see Algorithm 1) is cal-

led. The function consists of a loop with nested calcula-

tions of mean and variance. The worst-case time com-

plexity of that is

O((n − 2l

min

)(2n + 5) +

min

+ 2)

with

being size of the current window,

min

minimum

motion length and

constant. This can be simpliﬁed

O(n

)

. However, the window length is limited by

max

, so for a large number of data frames the com-

plexity of the segment function is not relevant. As a

result the overall time complexity for large

(with

: number of input frames) is

O(l

max

· c · n) → O (n)

which is necessary for an online algorithm working in

real-time.

4 EVALUATION OF THE

SEGMENTATION ALGORITHM

To evaluate the quality of the algorithm we use the re-

sults from the ofﬂine algorithms of (Zhou et al., 2013)

and (Krüger et al., 2016) as gold standard. Both pa-

pers used motion capture data of the publicly available

CMU database (Carnegie Mellon University, 2016), in

particular takes 01 to 14 from subject 86 (resampled

from

120 Hz

30 Hz

). We expect better results from

ofﬂine algorithms because they can exploit the full

temporal spectrum of the data, so they represent the

high bar of the results besides the ground truth. See

Figure 5 for a comparative overview of the results (our

algorithm with l

max

= 3.5).

The algorithms mentioned above differentiate bet-

ween motion segment and action class: the ﬁrst can be

seen as a more granular motion primitive whereas the

latter can be seen as a broader view to activities of a

human. The ground truth of the human observers only

refers to action classes and as our algorithm does not

recognize different action classes but simple segments,

we only compare to the ground truth’s action classes

and do not use the segment boundaries reported by the

other algorithms.

The output of the segmentation algorithms is a

time-series of segment boundaries. As the algorithms

may return time-series of different lengths (because

of different numbers of segments), we use Dynamic

Time Warping (DTW) (Müller, 2007) to calculate the

path with the minimal total cost that is required to

warp the time-series of an algorithm to the one of the

ground truth. The minimal total cost can be seen as

the minimal amount of frames that are required to shift

every segment boundary to a ground truth segment

boundary.

a time-series length

length

, then the

distance

DTW

of the optimal warping path

p∗

is (Mül-

ler, 2007):

DTW (X,Y ) = c

p∗

(X,Y ) = min

∑

l=1

c(x

, y

) (3)

In our case the cost function

c(x, y)

simply means

the distance

|x − y|

. Table 1 shows the normalized

DTW distances of our method and two state-of-the-art

algorithms to Ground Truth (GT) for every of the 14

trials.

HEALTHINF 2018 - 11th International Conference on Health Informatics

482

Trial 01

0 300 600 900 1200

Krüger et al. (2015)

0 300 600 900 1200

HACA

0 300 600 900 1200

Our method

0 300 600 900 1200

Frames

Ground Truth

Trial 06

0 500 1000 1500 2000 2500

Krüger et al. (2015)

0 500 1000 1500 2000 2500

HACA

0 500 1000 1500 2000 2500

Our method

0 500 1000 1500 2000 2500

Frames

Ground Truth

Figure 5: Graphical comparison of three segmentation algorithms with ground truth data for trials 1 (best case) and 6 (worst

case) of subject 86. Black bars mark the segmentation points. In case of (Krüger et al., 2016) the center of uncertainty was

chosen as the segmentation point.

5 DISCUSSION

The results show that the algorithm can continuously

segment motion data with robust although not supe-

rior quality compared to the ofﬂine segmentation al-

gorithms that can fully exploit the whole temporal

range of the data. In summary, the DTW distances

Unsupervised Temporal Segmentation of Skeletal Motion Data using Joint Distance Representation

483

Table 1: Normalized DTW distances of the three algorithms

to Ground Truth.

Trial

Our method

min

= 3.5)

(Krüger

et al., 2016)

HACA (Zhou

et al., 2013)

01 20.59 11.38 26.68

02 25.21 19.49 30.40

03 39.21 18.42 27.25

32.71 17.28 25.88

05 31.00 17.33 15.97

06 65.50 15.13 18.19

07 27.27 10.38 12.48

08 26.06 11.11 21.71

09 44.41 10.90 17.75

10 47.18 25.75 22.00

11 49.15 17.83 17.05

12 53.18 35.83 26.60

13 39.42 32.63 31.35

14 32.62 21.39 21.33

∅

38.04

frames

18.92 22.47

of our algorithms’ results are about twice as high as

of the reference algorithms, although the results vary

notably throughout the different trials. The quality of

a segmentation algorithm is of course highly depen-

dent on the tasks performed by the motion capture

subject. It is not surprising that the ofﬂine working

algorithms deliver results of higher quality, but on the

other hand, they have roughly quadratic complexity

and require a priori knowledge whereas our algorithm

has linear complexity and requires very little a pri-

ori knowledge (maximum window size). In practice,

the average DTW distance of our algorithm means

that the segment is 1-2 seconds away from the ground

truth, which is sufﬁcient if the motion segment is sig-

niﬁcantly longer. As a plus, our method returns the

motion dynamics variance of the segmented motion

which is used as static posture indicator, i.e., if the

algorithm returns a segment with low variance we as-

sume a static posture that can be further processed by

an ergonomic assessment method. In summary, the

method is a practically usable approach that is easy to

understand and implement although its accuracy falls

behind the state-of-the-art ofﬂine methods.

6 CONCLUSION

We described and evaluated an online-capable (real-

time) temporal segmentation algorithm for skeletal

motion data. The algorithm can be used to detect sta-

tic postures in a continuous stream of skeletal motion

capture data. Together with a digitalized ergonomic

assessment method such as OWAS (Karhu et al., 1977)

the detected postures can be used to derive risk fac-

tors for (work-related) musculoskeletal disorders. Due

to the online capabilities of the algorithm, it is pos-

sible to implement a live feedback system for users

of MoCap suits/systems when they perform unfavo-

rable or dangerous postures. (Yan et al., 2017; Ray

and Teizer, 2012; Peppoloni et al., 2014) are examples

for such systems and could possibly be used with our

algorithm. Because of its simple computability, the

algorithm works well on embedded hardware, since

only simple ﬂoating point calculations are necessary.

ACKNOWLEDGEMENTS

This work was partly funded by the German Ministry

for Education and Research (BMBF) within the joint

research projects SIRKA (grant 16SV6243), MeSiB

(grant 16SV7723) and by the funding initiative Nie-

dersächsisches Vorab of the Volkswagen Foundation

and the Ministry of Science and Culture of Lower

Saxony as a part of the Interdisciplinary Research

Centre on Critical Systems Engineering for Socio-

Technical Systems II. The authors would like to thank

Anna Vögele (Vögele et al., 2014) for kindly providing

the reference data of Table 1 and Figure 5.

REFERENCES

Aggarwal, J. K. and Cai, Q. (1997). Human motion analysis:

A review. In IEEE Nonrigid and Articulated Motion

Workshop Proceedings, pages 90–102, Washington,

DC, USA. IEEE, IEEE Computer Society.

Amell, T. and Kumar, S. (2001). Work-related musculoskele-

tal disorders: Design as a prevention strategy. a review.

Journal of Occupational Rehabilitation, 11(4):255–

265.

Armstrong, T., Buckle, P., Fine, L., Hagberg, M., Haring-

Sweeney, M., Martin, B., Punnett, L., Silverstein, B.,

Sjøgaard, G., Theorell, T., et al. (1996). Musculoskele-

tal disorders: Work-related risk factors and prevention.

International Journal of Occupational and Environ-

mental Health, 2(3):239–246.

Carnegie Mellon University (2016). CMU Graphics Lab

Motion Capture Database.

Corlett, E. N. (2003). Rapid Upper Limb Assessment

(RULA). Occupational Ergonomics: Principles of

Work Design, 1(June):9.

Coyte, P. C., Asche, C. V., Croxford, R., and Chan, B. (1998).

The economic cost of musculoskeletal disorders in

canada. Arthritis & Rheumatism, 11(5):315–325.

Ellegast, R. (2013). Gefährdungsbeurteilung am Arbeits-

platz. In Hartmann, Spallek, and Ellegast, editors,

HEALTHINF 2018 - 11th International Conference on Health Informatics

484

Arbeitsbezogene Muskel-Skelett-Erkrankungen: Ursa-

chen, Prävention, Ergonomie, Rehabilitation, pages

107–130. Hüthig Jehle Rehm.

Hignett, S. and McAtamney, L. (2000). Rapid Entire Body

Assessment (REBA).

Hoy, D., Brooks, P., Blyth, F., and Buchbinder, R. (2010).

The epidemiology of low back pain. Best practice &

research Clinical rheumatology, 24(6):769–781.

Karhu, O., Kansi, P., and Kuorinka, I. (1977). Correcting

working postures in industry: A practical method for

analysis. Applied Ergonomics, 8(4):199–201.

Keogh, E., Chu, S., Hart, D., and Pazzani, M. (2001). An

online algorithm for segmenting time series. In Pro-

ceedings 2001 IEEE International Conference on Data

Mining, pages 289–296. IEEE.

Koenig, N. and Matari

c, M. J. (2006). Behavior-based seg-

mentation of demonstrated tasks. In Proceedings of

the International Conference on Development and Le-

arning. Unknown.

Krüger, B., Vögele, A., Willig, T., Yao, A., Klein, R., and

Weber, A. (2016). Efﬁcient unsupervised temporal

segmentation of motion data. IEEE Transactions on

Multimedia, PP(99):1–1.

Kulic, D., Takano, W., and Nakamura, Y. (2009). Online

segmentation and clustering from continuous observa-

tion of whole body motions. IEEE Transactions on

Robotics, 25(5):1158–1166.

Lins, C., Eichelberg, M., Rölker-Denker, L., and Hein, A.

(2015). SIRKA: Sensoranzug zur individuellen Rück-

meldung körperlicher Aktivität. In 55. Wissenschaft-

liche Jahrestagung 2015 der Deutsche Gesellschaft

für Arbeitsmedizin und Umweltmedizin e.V., München,

pages 301–303. Deutsche Gesellschaft für Arbeitsme-

dizin und Umweltmedizin (DGAUM) e.V.

Lins, C., Müller, S. M., and Hein, A. (2016). Model-based

approach for posture and movement classiﬁcation in

working environments. In Wichert, R. and Klausing,

H., editors, Ambient Assisted Living: 8. AAL-Kongress

2015,Frankfurt/M, April 29-30. April, 2015, pages 25–

33. Springer International Publishing, Cham.

Matsui, H., Maeda, A., Tsuji, H., and Naruse, Y. (1997).

Risk indicators of low back pain among workers in

japan: association of familial and physical factors with

low back pain. Spine, 22(11):1242–1247.

Müller, M. (2007). Dynamic time warping. In Information

Retrieval for Music and Motion, pages 69–84. Springer

Berlin Heidelberg, Berlin, Heidelberg.

Nath, N. D., Akhavian, R., and Behzadan, A. H. (2017).

Ergonomic analysis of construction worker’s body pos-

tures using wearable mobile sensors. Applied Ergono-

mics, 62.

Peppoloni, L., Filippeschi, A., Ruffaldi, E., and Avizzano,

C. A. (2014). (WMSDs issue) A novel wearable sy-

stem for the online assessment of risk for biomechani-

cal load in repetitive efforts. International Journal of

Industrial Ergonomics, 52:1–11.

Punnett, L. and Wegman, D. H. (2004). Work-related muscu-

loskeletal disorders: the epidemiologic evidence and

the debate. Journal of Electromyography and Kinesio-

logy, 14(1):13–23.

Ray, S. J. and Teizer, J. (2012). Real-time construction wor-

ker posture analysis for ergonomics training. Advanced

Engineering Informatics, 26(2):439–455.

Roetenberg, D., Luinge, H., and Slycke, P. (2013). Xsens

MVN : Full 6DOF Human Motion Tracking Using

Miniature Inertial Sensors. Technical report, Xsens

Technologies B.V.

Schaub, K., Caragnano, G., Britzke, B., and Bruder, R.

(2013). The European Assembly Worksheet. The-

oretical Issues in Ergonomics Science, 14(6):616–639.

Stollenwerk, K., Vögele, A., Krüger, B., Hinkenjann, A., and

Klein, R. (2016). Automatic temporal segmentation of

articulated hand motion. In International Conference

on Computational Science and Its Applications, pages

433–449. Springer.

Vieira, A. W., Lewiner, T., Schwartz, W. R., and Campos,

M. (2012). Distance matrices as invariant features

for classifying mocap data. In Pattern Recognition

(ICPR), 2012 21st International Conference on, pages

2934–2937. IEEE.

Vögele, A., Krüger, B., and Klein, R. (2014). Efﬁcient

unsupervised temporal segmentation of human motion.

In Proceedings of the ACM SIGGRAPH/Eurographics

Symposium on Computer Animation, SCA ’14, pages

167–176, Aire-la-Ville, Switzerland, Switzerland. Eu-

rographics Association.

Walker, B., Muller, R., and Grant, W. (2003). Low back pain

in australian adults: the economic burden. Asia Paciﬁc

Journal of Public Health, 15(2):79–87.

Wang, D., Dai, F., and Ning, X. (2015). Risk assessment of

work-related musculoskeletal disorders in construction:

State-of-the-art review. Journal of Construction Engi-

neering and Management, 141(6):04015008.

Wang, L., Hu, W., and Tan, T. (2003). Recent develop-

ments in human motion analysis. Pattern Recognition,

36(3):585 – 601.

Yan, X., Li, H., Li, A. R., and Zhang, H. (2017). Weara-

ble IMU-based real-time motion warning system for

construction workers’ musculoskeletal disorders pre-

vention. Automation in Construction, 74:2–11.

Zhou, F., De la Torre, F., and Hodgins, J. K. (2008). Aligned

cluster analysis for temporal segmentation of human

motion. In Automatic Face & Gesture Recognition,

2008. FG’08. 8th IEEE International Conference on,

pages 1–7. IEEE.

Zhou, F., De la Torre, F., and Hodgins, J. K. (2013). Hierar-

chical aligned cluster analysis for temporal clustering

of human motion. IEEE Transactions on Pattern Ana-

lysis and Machine Intelligence, 35(3):582–596.

Unsupervised Temporal Segmentation of Skeletal Motion Data using Joint Distance Representation

485