ARTICLE IN PRESS

Journal of Biomechanics 42 (2009) 436442

Contents lists available at ScienceDirect

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

Motion segmentation method for hybrid characteristic on human motion

Newman Lau a,, Ben Wong a, Daniel Chow b
a b

Multimedia Innovation Centre, School of Design, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong Department of Health Technology and informatics, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong

a r t i c l e in f o
Article history: Accepted 29 November 2008 Keywords: Motion segmentation Repetitive motion Period identication Ergonomic evaluation

a b s t r a c t
Motion segmentation and analysis are used to improve the process of classication of motion and information gathered on repetitive or periodic characteristic. The classication result is useful for ergonomic and postural safety analysis, since repetitive motion is known to be related to certain musculoskeletal disorders. Past studies mainly focused on motion segmentation on particular motion characteristic with certain prior knowledge on static or periodic property of motion, which narrowed methods applicability. This paper attempts to introduce a method to tackle human joint motion without having prior knowledge. The motion is segmented by a two-pass algorithm. Recursive least square (RLS) is rstly used to estimate possible segments on the input human-motion set. Further, period identication and extra segmentation process are applied to produce meaningful segments. Each of the result segments is modeled by a damped harmonic model, with frequency, amplitude and duration produced as parameters for ergonomic evaluation and other human factor studies such as task safety evaluation and sport analysis. Experiments show that the method can handle periodic, random and mixed characteristics on human motion, which can also be extended to the usage in repetitive motion in workow and irregular periodic motion like sport movement. & 2009 Elsevier Ltd. All rights reserved.

1. Introduction The main purpose on ergonomic analysis is focused on quantied measurement on working posture, and how it relates to the occupational injuries or musculoskeletal disorder. Previous approaches required trained observers to give posture score based on some formulated methods (Latko et al., 1997, 1999; McAtamney and Corlett, 1993), which might induce large error and the result became subjective. To increase the accuracy and the quantity of measurement, automation process like video-based analysis was proposed to recognize and retrieve human-motion data (Albu et al., 2008; Blake et al., 1995; Gavrila and Davis, 1995; Lu et al., 2000; Min et al., 2008; Niyogi and Adelson, 1994; Yacoob and Black, 1998). Newer method like optical motion-tracking system was adopted by more researchers, especially in animation, movie industry and ergonomic analysis due to its simplicity and efciency in retrieving detailed and realistic human joint information. Though it had the shortcoming of occasional loss of data, solution was suggested by the previous study (Ormoneit et al., 2005). Ergonomic study requires analyzing on a specied joint or limb (a group of joints) given a gure on how injury is related to a particular motion or task. Some scoring methods were commonly

Corresponding author. Tel.: +852 2766 4345.

E-mail address: newman.lau@polyu.edu.hk (N. Lau). 0021-9290/$ - see front matter & 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.jbiomech.2008.11.038

used where the repetitiveness and static load could only be added on the overall score by the user subjectively (Hignett and McAtamney, 2000; McAtamney and Corlett, 1993). Those methods may not match with the modern needs of analyzing repetitive and static motion in quantitative and efcient way. Motion segmentation method was then proposed to represent motions as a sequence of movements or sub-motions, which were often modeled by dynamic systems. The method was discussed in numerous studies for not only repetitive motion analysis but also in dynamic manipulation, motion synthesis and robotic learning (Kimura et al., 2006; Troje, 2002). Different approaches were proposed like the statistical approach of principal component analysis (PCA) which represented the motion in a low-dimensional linear model (Jolliffe, 2002; Troje, 2002; Yacoob and Black, 1998; Zhang and Troje, 2007). Other researches also suggested modeling the data with dynamic non-linear systems (Kimura et al., 2006; Seitz and Dyer, 1997) or linear systems (Astrom, 1970; Astrom and Wittenmark, 1997; Blake et al., 1995; Lu et al., 2000; Lu and Ferrier, 2004). Event classication and motion-recognition techniques were also developed for switching and grouping the dynamic models for segmentation and further analysis (Blake et al., 1995; Campbell and Bobick, 1995; Gavrila and Davis, 1995; Lu and Ferrier, 2004; Min et al., 2008; Niyogi and Adelson, 1994; Yacoob and Black, 1998). The previous studies often assumed input motion as periodic, followed their focus on analyzing repetitive motion. However, in some cases, there are difculties in dening the composition of

ARTICLE IN PRESS
N. Lau et al., / Journal of Biomechanics 42 (2009) 436442 437

motion nature. We may not always have enough information to provide an accurate prediction on some joints movement like wrist twisting. Moreover, classifying a whole motion into either non-periodic or periodic is not reasonable, as the task may produce both repetitive and random complex motion by user variation and task nature. To prevent the analysis affected by subjective perception on specic motion and increase the generality of analysis, we should give no assumption on motion nature. This paper presented a two-pass method for gathering essential ergonomic data, which segmented and modeled human joint motion without the needs of any prior knowledge. The motion is preliminarily segmented into divisions and the loss of value of particular precision is guaranteed to be minimized (the rst-pass segmentation). The rst-passed divisions are then further segmented by motion nature (the second-pass). The nal segments are either representing a random motion or a cycle of periodic motion, which are modeled by damped harmonic model. This method can be applied to most of the motions, including periodic, random, or mixed motions. Many important motion parameters like motion range, frequency and duration can also be estimated. Experimental results demonstrated that different kinds of complex human motions could be handled by this method without stating any assumption. Discussion on the methods shortcomings, limitations and further improvements and developments were also presented at the later part of this paper.

where A 0 A2 I A1 ! and B 0 B0 !

Eq. (3) shows a rst-order auto-regression process in that A describes the deterministic part and B describes the stochastic part of the system. The segmentation problem is to nd where the dynamic tting parameters change signicantly. Parameter estimation of A(t) can be predicted by the recursive least square (RLS) model for time-varying system (Astrom, 1970; Ljung and Soderstrom, 1983) ^ ^ ^ At At 1 Ltwt wt 1At 1 Pt 1wt 1 Lt lt wT t 1Pt 1wt 1   1 Pt 1wt 1wT t 1Pt 1 Pt 1 Pt lt lt wT t 1Pt 1wt 1

(4)

Here, we present an enhanced segmentation process based on the previous study as the rst-pass algorithm (Lu and Ferrier, 2004): 1. Run a moving averaging on the input data to remove the noise of original motion-captured data. 2. Subtract the data by the mean to compact the data. 3. Apply recursive least square with discounted measurements to nd A1 and A2 at time t. 4. Compute the Frobenius norm of the difference matrix of A between current and previous time. 5. Find the turning points in the Frobenius norm graph. Mark the points as peak points if their value exceed a particular percentile value of the whole graph. 6. For each set of peak points found by particular percentile, set the value of all points to one and merge the nearby peak points by running a moving average. The moving-range should be sufcient to merge all scattered stems (peak points) into groups with single peak. In trial, the range is about 15% of the input motion length, according to complexity of motion. Smaller moving-range will produce segmentation points with higher precision. However, too short segment may not provide enough information for the model tting mentioned in the next section. 7. Set the highest points of each merged group as the possible segmentation points. 8. For each set of segmentation points, set A1 and A2 of each segment as constants by using the value of the starting time of that segment, which is estimated in Step 3. 9. A particular percentile value will produce a set of segmentation points. The optimized set of segmentation points is found by minimizing the sum of criterion (loss) function of each segment for all tested percentiles. If there are totally s segments, the loss function of each segment i 1, 2, y, s, each with starting time ti and length of Ni can be estimated by V i A
tXi i N kt i

2. Method 2.1. First-pass segmentation algorithm Joint angle is often adopted as a measure when researchers are dealing with ergonomic issues. Some papers divided joint angle into two- or three-dimensional space, especially for those video-based methods (Albu et al., 2008; Lu et al., 2000; Lu and Ferrier, 2004; Min et al., 2008; Niyogi and Adelson, 1994; Ormoneit et al., 2005; Polana and Nelson, 1997; Zhang and Troje, 2007). A hierarchical digital human model (DHM) was dened and the joint angle estimation was adopted according to the algorithm of hierarchical rotational matrix calculation mentioned in the previous study (Lau and Wong, 2007). Some researchers described human joint motion in multidimensional form in time series. For instance, upper arm, elbow and wrist are grouped as arm motion. In this paper, we would like to present the segmentation algorithm for analyzing localized risk factor on specic joint. Therefore, the motion is described as one-dimensional way to give detailed analysis on each joint motion. We treat the joint motion data (joint angle) as a sequence of signal without any prior knowledge of the characteristic. The sequence is then divided into meaningful segments. Let xtAR denotes the state vector of a joint (in this paper, the only state is the joint angle) at time t where t 1, 2, y, N. The auto-regression (AR) model is chosen to describe human motion since motion tracking requires prediction and measurement processes, in which a statistical framework of deterministic and stochastic model is adopted. And AR model, which relates previous data to new data, is similar to human-motion nature whose current state is related to future state (Ljung and Soderstrom, 1983): xt y jt vt
T

bNi ; kxk Ai1 xk1 Ai2 xk2 2

tk

(5)

where Ai and Ai are the values of A1 and A2 of segment i, bt; k l by 1 2 supposing the forgetting factor l(k) l for all ks and we use l 0.98 in the experiments.

(1)

where yT (A1, A2, y, AN) are the parameters for describing the dynamic relationship between the signal with its previous value. jT(t) (xt1, xt2, y, xtN) represents the lagged output data, v(t) states for the disturbance, which is set as a zero mean white noise. In this paper, second-order auto-regression model is used to represent the motion data, which can describe complicated human-motion dynamics (Lu et al., 2000): xt A1 xt1 A2 xt2 vt (2)

where A1 and A2 are dynamic parameters representing the motion. Eq. (2) can be written in a more compact form by using a state vector w(t) (xt1x, xt2x)T where x is the mean of input data, and converting v(t) into the following form (Lu et al., 2000):

wt Awt 1 Bet

(3)

The Frobenius norm of difference matrix of A actually shows the change of tting parameters (A1 and A2) in the motion. Turning point means a possible change of tting model in the dynamic system, thus possible segmentation point. The accuracy of the algorithm depends on how precise is it in choosing the percentile level. In the experiment, the percentile level was tested from 99 down to 50 with one decrement. The only parameter for this process is the moving-range in Step 6 (suppose enough large percentile range and suitable precision are used). Smaller values produce more segmentation points, with smaller loss, suitable for random motion, and larger values produce fewer segmentation points, dividing motion into groups of periodic motion with multiple cycles, suitable for low-varying or periodic data. A general example of the lower arm joint angle of a hand-waving motion with 20 frame per second (FPS) is shown in Fig. 1 with higher precision and Fig. 2 with lower precision. The segmentation with larger moving-range included a group of periodic data, while that in smaller moving-range segmented motion into nearly half cycle for a segment. In practice, control for the output using smaller movingrange is difcult. Especially in low-varying or periodic motion, small moving-range may segment motion into either single cycles or half cycles. In this paper, state vector in RLS is one-dimensional only. If the dimension is increased by including multiple joint angles inside the state vector xk. The above algorithm would still applicable by changing the matrix dimension of tting parameters and the rst-passed Frobenius norm should act as a new motion data and repeat Steps 14 before going into Step 5 if this one-pass algorithm does not give satisfactory result.

ARTICLE IN PRESS
438 N. Lau et al., / Journal of Biomechanics 42 (2009) 436442

0.4 0.3 80 0.2 50 55 60 65 70 75 80 85 90 95 100 60 0.04 0.02 0 0 1 0.5 0 0 20 40 60 80 100 120 140 160 180 200 40 60 0 50 100 150 200 250 20 20 40 60 80 100 120 140 160 180 200 0 20 40

2 0 2 0 20 40 60 80 100 120 140 160 180 200

Fig. 1. (a) Criterion (loss) value in RLS segmentation method with square-block locating the minimum loss value, (b) Frobenius norm of the difference matrix A, (c) peak points on (b) where the values are larger than the optimized percentile found in (a). Moving averaging with range of 2% is executed to produce separate groups of stem, (d) rst-pass segmentation result where each step states a segmentation point and (e) input motion signal of the lower arm joint angle on a hand-waving motion at 20 FPS and segmentation points.

(e) 1 0.5 60 0 50 55 60 65 70 75 80 85 90 95 100 40 80

0.04 0.02 0 0 20 40 60 80 100 120 140 160 180 200

20

20 1 0.5 0 0 20 40 60 80 100 120 140 160 180 200 60 0 2 0 2 0 20 40 60 80 100 120 140 160 180 200 50 100 150 200 250 40

Fig. 2. Same as Fig. 1 with changed moving-range to 5%. which can be found by the eigenvalues of parameter matrix A, l (Blake et al., 1995) b i2pf 1 log l (7)

2.2. Second-pass segmentation algorithm and model tting Previous work used a damped harmonic model to t on each segment of motion (Lu and Ferrier, 2004)

a expbt cos2pft f T

(6)

where a, b, f, f and T represent the amplitude, damping rate, frequency, phase and offset information of the damped harmonic oscillation, respectively. Those parameters could be estimated by subjecting to a non-linear least square t by the LevenbergMarquardt method (Hiremath, 2006; Press et al., 1992). Frequency and damping rate in Eq. (6) should be bounded into positive value,

where D is the sampling rate. Initial value of the amplitude is set as the mean of the segment data. Starting values of phase and offset are zero. As the damped factor in the model will inuence the motion in a relatively short time, this model is weak to simulate multiple-cycle motion. If the movingrange mentioned in the previous section produces low-precision result which contains multiple cycles of motion in a segment, the model tting may produce unsatised result or fail.

ARTICLE IN PRESS
N. Lau et al., / Journal of Biomechanics 42 (2009) 436442 In practical, the sub-motions may only consist of a part of the mentioned damped harmonic motion. In such cases, the amplitude of the model may be too large for representing the amplitude of motion, which is the actual varying range in the observed data. Thus, the actual motion amplitude should be the smaller one of the varying range in that segment and the estimated a, i.e. min[maxdiff(segment), a], where maxdiff(X) is the largest difference value in data X, that is the varying range. In experiments, we set a medium moving-range of about 2% of the data length, which is a low-precision setting in most cases, assuming that we have no prior knowledge on the motion characteristic. Then we will encounter two problems, one is that multiple-cycle motion may occur that fails the tting, and a singledamped harmonic may not well t with the random movement in a long segment. In this paper, we adopted an identication method described in the work of Ormoneit et al. (2005), which was originally used for period identication, as the second-pass segmentation method. For each segment i with length Ni, the error of damped harmonic tting could be determined as the normalized root-meansquare deviation (NRMSD), which is represented in percentage q P ^ 1=N i ti Ni xk xk 2 kti maxxti ; xti Ni minxti ; xti Ni 439

3. Results Experiments were done for demonstration and testing the functionality of the presented method. An example motion of angle on right wrist-twisting in a hand-waving motion tested at 20 FPS is shown in Fig. 4. Though we knew the motion might be periodic, we assumed no prior knowledge. The above algorithm was applied to produce the tting result as shown in Fig. 5 with parameters (frequency, amplitude and duration) summarized in Table 1. Result showed that the motion appeared to be periodic but with irregular periods and amplitude, which was often happened in daily human motion. With the rst-passed segmentation points, the damped harmonic model could not give a satised result (Fig. 5a). Together with the extra segments (second-passed), the result could be well-tted on this kind of motion characteristic (Fig. 5b). Parameters show relatively high-frequency motions which may relate to repetitive movement. Signicant range of joint
20 15 10 5 0 5 10 15 20 0 50 100 150 200 250

Ei

(8)

where ti states the starting time of segment i and xk are the tted values by damped harmonic model at time k. If Ei exceeds certain threshold Ti, the segment may contain multiple-cycle or is a high-complexity motion. In the experiment, setting a constant T 3% for all segments is always satisfactory. We should then apply the method and estimate the most possible period to set extra segment points. Each sub-segment is then tted by a damped harmonic model individually. This period identication method tests for a range of predicted period (bounded by the segment length), and trials will be performed to test the signalto-noise ratio (SNR) of the clipped motion. High SNR demonstrates possible period value. However, if the predicted period range is too large, it may contain harmonics of the signal, caused multiple high SNR values, confusing on how to identify the true period. Further studies would be required to give a pre-analysis on the input motion to provide possible range of period. In this paper, the range is bounded to half of the signal length that shows validity on most of the examples and experiments resulting only single-peak SNR. Combining with the segmentation method, the whole analysis produces a twopass algorithm which is useful to model large variation of motion characteristic. In the experiment, large moving-range was used to set the result into low precision, suitable for motion of no prior knowledge. The error threshold T will act as an indicator to locate where extra segmentation should be performed to model the local characteristic. Fig. 3 shows the model tting of the same data set and moving-range as in Fig. 2. With the aid of the second-pass process, motion could be correctly segmented into individual cycles and meaningful segments. The tting method also provides frequency, amplitude and motion duration, which can be applied for evaluation on ergonomics and occupational health, in which the frequency and amplitude greatly affect the repetitive stress on joints and static work load.

Fig. 4. Right-wrist-twisting angle in a hand-waving motion at 20 FPS.

60 40 20 0 20 40 0 50 100 150 200 250 original data fitted data segment points

60 40 20 0 20 40 0 50 100 150 200 250 original data fitted data segment points sub segment points

Fig. 3. (a) Damped harmonic tting without using the second-pass segmentation algorithm and (b) model tting with further segmentation (the second-pass process).

ARTICLE IN PRESS
440 N. Lau et al., / Journal of Biomechanics 42 (2009) 436442

20 10 0 10 20 0 20 10 0 10 20 0 50 100 150
original data fitted data segment points sub segment points original data fitted data segment points

50

100

150

200

250

200

250

Fig. 5. Segmentation result of motion shown in Fig. 4: (a) joint angle, damped harmonic tting and segmentation points without using second-pass segmentation and (b) same as (a) with extra segmentation points.

Table 1 Frequency, amplitude and duration of right wrist-twisting angle in hand-waving motion. Segment 1 2 3 4 5 6 7 Frequency (Hz) 0.23 0.18 0.77 0.66 0.58 0.55 0.47 Amplitude (deg.) 0.08 8.04 5.38 6.54 9.92 5.86 8.38 Duration (s) 2.50 1.10 0.65 2.35 1.40 1.40 0.65

variation of researches such as the sport motion analysis, and task-safety evaluation under ergonomic considerations.

4. Computational complexity Suppose an input signal has length of N. The process of RLS algorithm and the process of minimizing the loss function in the rst-pass algorithm have computational complexity of O(N2) and O(N), respectively, bounded the complexity to O(N2) for the rstpass. In the second-pass algorithm, the efciency of computation on the damped harmonic model by non-linear least square t using the LevenbergMarquardt method depends on the sample size and the initial guess of coefcients. A reasonable initial guess can lead to fast convergence of the iteration process. More studies need to be done on how to estimate good initial coefcients for the process. Computation of NRMSD is rather simple and has a O(N) complexity. For each of the segment from the rst-pass algorithm with length n, the maximum number of computation of SNRs in the period identication method is n4/2, if the period is tested from 1 to n/2, resulting computation complexity bounded by O(N4). Theoretically, the minimum length of a segment represented by second-order autoregressive process is three. However, in such situation, the period identication method cannot be applied. In contrary, the maximum length of a segment is N, which turns the computation complexity of period identication into O(N4). For other segment length, the number of segment in the input data is relatively insignicant to the segment length n and input data length N. Then, we know that the number of segments will not affect the computation complexity and come to a conclusion that the computation complexity of the period identication is bounded by O(N4), as N4n. The period-identication method can be treated as a brute-force process to nd the period inside a segment, thus producing such high computation complexity. All in all, the computation complexity of the whole proposed method is bounded by O(N4).

motion may also link to occupational and musculoskeletal risk factors (Buckle and Devereux, 1999; Grandjean, 1988; Latko et al., 1999; McAtamney and Corlett, 1993). However, short motion time, of about 10 s, may not produce risk to human. Motion characteristic in this example also appears in many sport movements like kicking the football with different strength. Another example of right upper arm movement in a desktop work is illustrated in Fig. 6 at 1 FPS. As no prior knowledge on the working nature was given, large moving-range was used. The ergonomic parameters are listed in Table 2. The data consisted of random motion in the rst half and partly periodic motion in the second half. As low-precision moving-range was used, the rst half could not be modeled by a single-harmonic model (Fig. 7a). Thus, extra points were provided for effective modeling, and the simulation in the periodic part was also improved (Fig. 7b). Frequencies in this motion are relatively low and motion durations are long, which may link to static work problem (Grandjean, 1988; McAtamney and Corlett, 1993). This example can demonstrate the ability of the method to model the hybrid characteristic of motion and applicability for motion evaluation for long task. The experiments show that for different kinds of motion characteristics, the two-pass algorithm can effectively divide the human motion and model the result segments. The examples can also demonstrate the applicability of this method in large

ARTICLE IN PRESS
N. Lau et al., / Journal of Biomechanics 42 (2009) 436442 441

50

45

40

35

30

25

20

15 0 50 100 150 200 250 300 350 400

Fig. 6. Joint angle of right upper arm in a desktop task at 1 FPS.

Table 2 Frequency, amplitude and duration of right upper arm in desktop work. Segment 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Frequency (Hz) 0.03 0.04 0.04 0.04 0.04 0.03 0.06 0.05 0.03 0.07 0.03 0.03 0.03 0.03 Amplitude (deg.) 26.21 2.72 6.58 6.74 0.58 1.78 2.31 4.88 2.12 0.52 4.76 7.24 5.79 8.20 Duration (s) 23.00 23.00 23.00 23.00 23.00 23.00 23.00 23.00 23.00 18.00 16.00 48.00 48.00 24.00

5. Conclusion This paper presented a method to automatically segment complex motion data without any prior knowledge on motion characteristic, and simulate each individual segment with a damped harmonic model. This routine shows effectiveness on segmenting different motion characteristics into reasonable segments and performing desirable model tting. With relatively fewer control parameters, only the moving-range and the error threshold T, the control of this segmentation routine is simple and easy. This method could also be extended for event identication (Blake et al., 1995; Campbell and Bobick, 1995; Gavrila and Davis, 1995; Lu and Ferrier, 2004; Min et al., 2008; Niyogi and Adelson, 1994; Yacoob and Black, 1998) and enhanced into high dimensional, providing exibility and extensity for different research aspects.

There are also some drawbacks in the method presented here. Firstly, user has to consider running a low-pass lter with different settings according to the input-data quality for removing the data noise. Secondly, the error threshold needs be tuned to prevent or under-segmentation in the second-pass algorithm. User may refer to our experimental value 3% as the starting point of the error threshold. Lastly, the main purpose of this paper is to propose a highly applicable algorithm for motion segmentation, which provides high-quality result and is easy to control. Therefore, the internal computation processes are not optimized. Implementation of improved recursive least squares, damped harmonic t and period identication processes left for further studies. An example of alternative of the periodidentication method can also be found in the wok of Albu et al. (2008). As this paper aims to provide the method that may segment the motion to provide essential data for further analysis on ergonomic issues, analysis given in the experimental results is short and rough. User should also be aware that the amplitudes given by the method (in Tables 1 and 2) are only the varying range of joint angle of that segment, but not the actual value of jointangle movement, as the data are subtracted by their mean and being low-passed.

Conict of interest statement We (Newman Lau and Ben Wong) hereby warrant that we do not have any nancial and personal relationships with other people or organizations that could inappropriately inuence (bias) our work in producing the paper with title of Segmentation Method for Hybrid Characteristic on Human Motion.

ARTICLE IN PRESS
442 N. Lau et al., / Journal of Biomechanics 42 (2009) 436442

20 10 0 10 20 30 40 0 10 0 10 20 30 40 0 50 100 150 200 250 300 350 400 original data fitted data segment points sub segment points 50 100 150 200 250 300 350 400 original data fitted data segment points

Fig. 7. Segmentation result of motion shown in Fig. 6: (a) joint angle, damped harmonic tting and segmentation points without using second-pass segmentation and (b) same as (a) with extra segmentation points.

Acknowledgments The authors thank Leung, Wing Sang for his technical assistance on motion capture processes. This research is funded by the Departmental General Research Funds of the Hong Kong Polytechnic University. References
Albu, A.B., Bergevin, R., Quirion, S., 2008. Generic temporal segmentation of cyclic human motion. Pattern Recognition 41, 621. Astrom, K.J., 1970. Introduction to Stochastic Control Theory. Academic Press, New York. Astrom, K.J., Wittenmark, B., 1997. Computer Controlled SystemsTheory and Design. Prentice Hall Information and System Sciences Series, Englewood Cliffs, NJ. Blake, A., Isard, M., Reynard, D., 1995. Learning to track the visual motion of contours. Articial Intelligence 78, 179212. Buckle, P., Devereux, J., 1999. Work-related neck and upper limb musculoskeletal disorders. Ofce for Ofcial Publications of the European Communities. Campbell, L.W., Bobick, A.F., 1995. Recognition of human body motion using phase space constraints. In: Proceedings of the Fifth International Conference on Computer Vision. Gavrila, D.M., Davis, L.S., 1995. Towards 3-D model-based tracking and recognition of human movement: a multi-view approach. International Workshop on Face and Gesture Recognition, pp. 272277, /http://citeseerx.ist.psu.edu/viewdoc/ summary?doi=10.1.1.56.5329S. Grandjean, E., 1988. Fitting the Task to the Man, A textbook of Occupational Ergonomics, fourth ed. Taylor & Francis, London. Hignett, S., McAtamney, L., 2000. Rapid entire body assessment (REBA). Applied Ergonomics 31, 201205. Hiremath, K.M., 2006. The solar cycle as a forced and damped harmonic oscillator: long-term variations of the amplitudes, frequencies and phases. Astronomy and Astrophysics 452, 591596. Jolliffe, I.T., 2002. Principal Component Analysis, second ed. Springer, Berlin (Series in Statistics). Kimura, H., Tsuchiya, K., Ishiguro, A., Witte, H., 2006. Dynamic Movement PrimitivesA Framework for Motor Control in Humans and Humanoid Robotics. Springer Tokyo, Berlin, pp. 261280.

Latko, W.A., Armstrong, T.J., Foulke, J.A., Herrin, G.D., Rabourn, R.A., Ulin, S.S., 1997. Development and evaluation of an observational method for assessing repetition in hand tasks. American Industrial Hygiene Association Journal 58, 278285. Latko, W.A., Armstrong, T.J., Franzblau, A., Ulin, S.S., Werner, R.A., Albers, J.W., 1999. Cross-sectional study of the relationship between repetitive work and the prevalence of upper limb musculoskeletal disorders. American Journal of Industrial Medicine 36, 248259. Lau, N., Wong, B., 2007. Engaging rapid user-centered assessment in biomechanics and ergonomic at early product design phase. In: Proceedings of the Third International Conference on Product Design and Development. Ljung, L., Soderstrom, T., 1983. Theory and Practice of Recursive Identication. The MIT Press, Cambridge, MA. Lu, C.M., Ferrier, N.J., 2004. Repetitive motion analysis: segmentation and event classication. IEEE Transactions Pattern Analysis and Machine Intelligence 26, 258263. Lu, C.M., Liu, H., Ferrier, N.J., 2000. Multidimensional motion segmentation and identication. IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) 02, 2629. McAtamney, L., Corlett, E.N., 1993. RULA: a survey method for the investigation of work-related upper limb disorders. Applied Ergonomics 24, 9199. Min, J., Kasturi, R., Camps, O., 2008. Extraction and temporal segmentation of multiple motion trajectories in human motion. Image and Vision Computing 26, 16211635. Niyogi, S.A., Adelson, E.H., 1994. Analyzing and recognizing walking gures in XYT. In: Proceedings of Computer Vision and Pattern Recognition (CVPR), IEEE Computer Society Conference, /http://ieeexplore.ieee.org/xpls/abs_all.jsp? arnumber=323868S. Ormoneit, D., Black, M.J., Hastie, T., Kjellstrom, H., 2005. Representing cyclic human motion using functional analysis. Image and Vision Computing 23, 12641276. Polana, R., Nelson, R.C., 1997. Detection and recognition of periodic, nonrigid motion. International Journal of Computer Vision 23, 261282. Press, W.H., Flannery, B.P., Teukolsky, S.A., Vetterling, W.T., 1992. Numerical Recipes in C: The Art of Scientic Computing, second ed. Cambridge University Press, Cambridge. Seitz, S.M., Dyer, C.R., 1997. View-invariant analysis of cyclic motion. International Journal of Computer Vision 25, 231251. Troje, N.F., 2002. Decomposing biological motion: a framework for analysis and synthesis of human gait patterns. Journal of Vision 2, 371387. Yacoob, Y., Black, M.J., 1998. Parameterized modeling and recognition of activities. In: The Proceedings of Sixth International Conference on Computer Vision. Zhang, Z., Troje, N.F., 2007. 3D periodic human motion reconstruction from 2D motion sequences. Neural Computation 19, 14001421.