MODELING AND TRANSFORMATION OF 3D HUMAN MOTION

Seyed Ali Etemad

Department of Systems and Computer Engineering, Carleton University, 1125 Colonel By Drive, Ottawa, Canada

Ali Arya

School of Information Technology, Carleton University, 1125 Colonel By Drive, Ottawa, Canada

Keywords: Piece-wise Time Warping, Synthesis, Style Transformation, Critical Features.

Abstract: Applying different styles of motion such as those related to gender, age and energy of the performer, are

important themes in creating realistic animation. While typically different sets of motion sequences must be

created for different styles, in this paper we present a technique for transferring motion styles across

different sequences. This allows transforming the style of a motion while preserving the primary action

class of the original motion. A mathematical model capable of defining both the action class and the style

class is proposed and based on the model, the conversion data are computed. A series of piece-wise time

warping procedures are conducted prior to employing the defined transformation function. Using this

technique, there would be no need for a new set of motion data to be captured or keyframed as the original

motion data can be altered to show the desired style.

1 INTRODUCTION

The use of optical motion capture systems to

produce high quality humanoid motion sequences

has recently become increasingly popular in many

areas such as high-end gaming, movie and animation

production, and for educational purposes. Motion

capture studios are expensive to employ or are

unavailable to many people and locales. For this

reason, the reuse of existing motion capture data and

the resulting high-quality animation is an important

area of research. Also strict motion capture data

restricts users to pre-recorded movement that does

not allow the addition of dynamic behaviours

required in advanced games and interactive

environment.

Procedural animation is creation of motion

sequences through an algorithm rather than

conventional methods like keyframing and motion

capture. Creating new styles (i.e. personal variation

of a base action) of animation based on existing

motion sequences is an example and a relatively new

topic of research in the field of animation where the

goal is to manipulate existing sequences to create

animation of the same action class, yet in a different

mode. For instance using such techniques, a walking

action sequence can be transformed from its initial

masculine mode to a feminine one. Similarly the age

and energy level of the characters can be controlled

and the respective motion data can be synthesized.

The goal of this research is to create a system

capable of learning various action classes with

specific styles related to its character, and transform

the style based on the user desire. Examples of the

basic classes of action are the three periodic actions

of walking, jumping and running. While walking

and jumping are considerably different, walking and

running are relatively similar, thus synthesis of

motion data for these actions must be especially

accurate.

For each basic class of action, secondary

variation style pairs are necessary for adding a level

of personality to the motions. Articles relating to

personality and behaviour such as (Sok, et al., 2007;

Su, et al., 2005; Kang & Kim 2007) were reviewed

in order to determine contributing factors of human

locomotion that can show a character’s personality.

These personality variations add a level of

personality and interest to the basic, unexaggerated

animation types mentioned earlier. Masculine to

Feminine, Old to Young, Tired to Energetic, Happy

to Sad, and Determined to Dreamy are these

307

Ali Etemad S. and Arya A. (2010).

MODELING AND TRANSFORMATION OF 3D HUMAN MOTION.

In Proceedings of the International Conference on Computer Graphics Theory and Applications, pages 307-315

DOI: 10.5220/0002755003070315

 SciTePress

animation variation styles. By incorporating the

variations with the basic animation types, new and

unique character animations can be produced. The

presented procedural animation method is a proof of

concept production, and as such, only a selection of

the animation types and variations are tested using

our proposed method. The Walk, Run, and Jump

animation types with the Masculine to Feminine,

Old to Young, and Tired to Energetic variations are

included. Yet the system can be extended to include

other styles in the future.

We have proposed a novel mathematical model

for describing human actions with various styles; the

same model has been used to describe each step. The

proposed transformation method is based on

applying a transfer function obtained during a

training phase to the base motion sequence in order

to create a desired motion. The first step for style

transformation of actions is for the training data to

be temporally equal in length i.e. motion data

matrices must have same number of frames in order

for us to perform mathematical operations on them.

We have proposed a novel piece wise time warping

technique to convert our motion data sets to data sets

of the same temporal length. The model has then

been used to generate the necessary transfer

functions for style transformations between different

styles of the same action.

To test the outputs of our transformation

techniques we formed a questionnaire and arranged

for participants to exercise various transformations

applied on different actions. The user comments and

ratings confirm the significance of our research. Our

procedural animation method offers animators high

quality animations produced from an optical motion

capture session, without incurring the cost of

running their own sessions. This method utilizes a

database of common animations sequences, derived

from several motion capture sessions, which

animators can manipulate and apply to their own

existing characters through the use of our procedural

animation technique.

2 RELATED WORK

In recent years, much research has been

conducted with the aim of synthesizing human

motion sequences. Statistical models have been one

of the practical tools for human motion synthesis

(Tanco & Hilton, 2000; Li, et al., 2002). Tanco and

Hilton (2000, pp. 137-142) have trained a statistical

model which employs a database of motion capture

data for synthesizing realistic motion sequences and

using the start and end of existing keyframes,

original motion data are produced. Li et al. (2002,

pp. 465-472) define a motion texture as a set of

textons and their distribution values provided in a

distribution matrix. The motion texton is modeled by

a linear dynamic system (LDS). A maximum

likelihood algorithm is designed to learn from a set

of motion capture based textons. Finally, the learnt

motion textures have been used to interactively edit

motion sequences.

Egges et al. (2004, pp. 121-130) have employed

principal component analysis (PCA) to synthesize

human motion with the two deviations of small

posture variations and change of balance. This

approach is useful in cases where an animated

character is in a stop/freeze situation where in reality

no motionless character exists. Liu and Papovic

(2002, pp. 408-416) have applied linear and angular

momentum constraints to avoid computing muscle

forces of the body for simple and rapid synthesis of

human motion. Creating complex dynamic motion

samples such as swinging and leaping have been

carried out by Fang and Pollard (2003, pp. 417-426)

using an optimization techniques applied along with

a set of constraints, minimizing the objective

function. Pullen and Bregler (2002, pp. 501-508)

have trained a system that is capable of synthesizing

motion sequences based on the key frames selected

by the user. Their method employs the characteristic

of correlation between different joint values to create

the missing frames. In the end quadratic fit has been

used to smooth the estimated values, resulting in

more realistic looking results. Brand and Hertzmann

(2000, pp. 183-192) employ probabilistic models for

interpolation and extrapolation of different styles for

synthesis of new stylistic dance sequences using a

cross-entropy optimization structure which enables

their style machine to learn from various style

examples. Safonova et al. (2004, pp. 514-521) define

an optimization problem for reducing the

dimensionality of the feature space of a motion

capture database, resulting in specific features.

These features are then used to synthesize various

motion sequences such as walk, run, jump and even

several flips. This research shows that the complete

feature space is not required for synthesis of human

motion. We have employed this property in section 5

where correlated joints have been ignored when

transforming the actor style themes.

Hsu et al. (2005, pp. 1082-1089) conduct style

translations such as sideways walk and crouching

walk based on a series of alignment mappings

followed by space warping techniques using an LTI

model. While this technique shows to be functional

GRAPP 2010 - International Conference on Computer Graphics Theory and Applications

308

for the mentioned style translations, more minute

style variations such as those related to gender,

energy and age have not been tested for. Rose et al.

(1998, pp. 32-40) have employed time warping as

the first step towards synthesis of human motion

which is an approach similar to ours. Specific

kinematic constraints have been exercised along

with interpolation of styles. The constraints,

however, are selected manually and based on the

nature of the action, as opposed to our proposed

method in chapter 6 where we tend to automate the

constraints. Time warping is also studied by Hsu et

al. (2007, pp. 45-52) where they have used a specific

reference motion for defining the constraints of the

time warping procedure. Although this technique

proves to be useful in some cases, it is not discussed

whether the reference action is capable of

determining the most suitable constraints for all

cases.

3 DATASETS AND MOTION

THEORY

Motion capture data obtained by means of a six

MX40 camera Vicon system based in Carleton

University have been used for this research. We

have asked various actors to perform the required

basic class of actions in each variation style, and the

necessary database has been created.

The motion capture data come in the form of (1),

where Di are the Cartesian values for the hip marker

in 3D space with respect to the calibration origin and

are the rotation angles in degrees for each marker.

There are m rows, denoting m frames.



















(1)

In more details, the motion data matrices are in

the form (2) where



represents the rotation values

of the x

axis of the j

marker for the i

frame and

represents the position of the x

axis of the hip

marker for the i

frame.

















nnn

zyx

ddd



,,,...,,,

111

222222

111111

222

111



(2)

The frame rate of A is consistent throughout the

whole matrix and the time lapse between each two

consecutive frames is fixed and equal to 0.05

seconds. It is very important that all motion samples

be of the same frame rate.

For this research, two different sets of data have

been acquired. The first data set is used for training

the system and creating the transformation values,

while the second is used for testing the system as

will be discussed later on in section 6.

Human action and motion in general is a

combination of an action class and a set of stylistic

variations adjoined to it. We call the action classes

which are the dominant signals throughout the data

sets, primary themes, and the stylistic variations as

secondary motor themes (Hutchinson, 1996). Based

on this definition, any action can be modeled by (3)

where Y[k,r] is the action sequence with the primary

theme k and secondary theme r as it is observed,

P[k] is the primary motor theme of the same action

class, and S[r] is the secondary theme of class r. In

(3) w[r] is the weight applied to the secondary

themes, and e represents the noise available in the

model.

  

erSrwkPfrkY





1

][][][]:1,[

(3)

The model (3) is defined such that a combination

of different styles can be applied to the same

primary class of action. For instance, for action class

jump, the secondary theme young-feminine can be

defined. A total of f different S functions are

foreseen in the model. Yet while training the system,

all samples are designed and assumed to hold only

one secondary theme. Each secondary theme

function is learnt separately, and ultimately, they can

be merged to form a multi-style secondary theme.

The goal of this research is to create the S[r]

values, while trying to minimize e. In most cases,

the S[r] signals are not powerful enough to influence

the perception of k, i.e. the action class. But rarely,

this scenario might take place. For instance if r

relates to class of energy and speed related themes,

S[r] along with a notable w[r] applied to a primary

theme of walking can cause confusion as to if Y[k,r]

was originally walking with a strong S[r] and w[r],

or whether k determines the action class of running

for Y[k,r], and S[r] and w[r] have been insignificant.

This dilemma is addressed in section 6, yet further

research is ongoing to tackle this issue.

MODELING AND TRANSFORMATION OF 3D HUMAN MOTION

309

4 FEATURE SELECTION AND

PIECEWISE TIME WARPING

The captured motion data matrix A can be in various

lengths. While some actions have been performed

faster or slower due to actor preferences, some

actions are by nature faster or slower than others.

For instance, a 2-step walk cycle takes more time

compared to a 2-step run cycle. In order to

manipulate such data matrices based on one another,

the temporal length of each action should be warped

to match the temporal length of the relative actions

with which we intend to blend. As for the number of

columns in the matrices, they do not require any

manipulation since the exact same number and

orientation of markers have been used for all capture

sessions.

The system is trained based on the original input

sequence and the target sequence. The first set is the

base animation while the second is the target

sequence containing the desired secondary theme.

Together they are called the training data. The

system is then tested for new sets of action

sequences called the test data. The training data

along with the test sequence are temporally warped

such that they are all in length. A straightforward

approach is simple scaling; that is stretching or

shrinking the motion sequences. Yet the sequences

must be aligned appropriately. To tackle this

problem, the warping of the matrices is carried out

with the goal of aligning some critical features of the

three datasets.

To conduct the very crucial piece-wise time

warping, the signals are divided into two sections.

The corresponding first sections are warped

together, and the second parts are warped together as

well. This is carried out in order for different

corresponding sections of the action sequences to be

aligned correctly. Thus, selecting the proper feature

for which the signals are sliced at is very critical.

Different methods were tested for finding the

critical features, used to determine the boundaries

for the piece-wise time warping. The first method is

by manually selecting the features. For walk and run

sequences, the feature instance was selected as the

occurrence of a foot touching the ground. This

means in a sequence where walking starts with the

right foot initiating the walk and ending in the same

situation, the instance when the left foot touches the

ground is labelled as the feature. For the jump

sequence, the feature is defined as the instance

where the actor is at the peak of his/her jump (when

the inclining motion stops and returning to the

ground initiates). Figures 1 and 2 illustrate the

piecewise warping conducted using this technique.

In Figure 1 through Figure 5, the data for the x axis

of the marker placed on the right leg are plotted.

Figure 1: Signals from training, target, and test matrices

before time warping.

It can be noticed from Figure 2 that the warped

signals are stretched to last the exact amount of time

(46 frames) which is the average of each of the

individual signals in Figure 1. Yet misalignment in

the global minimum peaks has occurred. Also the

alignment of the signals between frame 20 and 30 is

disposed of which is an undesired artefact. Thus

selecting the features manually is not recommended.

Figure 2: Signals from training, target, and test matrices

after time warping using manual feature selection.

The second method is by means of statistical

analysis. The A matrices are first normalized. The

normalized version of A is called A

. A matrices are

then differentiated with respect to time as presented

by (4) where the result is denoted by



presented by

(5) and (6).







(4)

GRAPP 2010 - International Conference on Computer Graphics Theory and Applications

310





















nnnnnn

nnnnnnn

zzyyxx

zzyyxxx

zyyxx



111111

323232323232

212121212121

,,,...,,,

111111

11111





(5)















mmm

vvv

321

,,,









(6)

The vector Φ is then calculated by (7).





















(7)

The Φ vector determines the net velocity of all

markers for each frame. The total velocity of all

markers is a very determinant informative value.

The maximum or minimum arrays of Φ verify the

instances where the actor has reached his/her

maximum angular velocity of the action. The largest

column value in Φ determines the critical feature,

which is where the maximum angular velocity is

observed. Assuming that during the course of an

action, the markers accelerate from zero angular

velocity to this value and then accelerate back to

zero, time warping is carried out for each section of

the action. The result is actions of the same length

and temporally aligned such that sections of the

action with positive angular acceleration and

negative angular acceleration are aligned. Both the

maximum and minimum values were tested. Figures

3 and 4 show the effect of selecting the features

based on this method.

It can be noticed that in Figure 3 and Figure 4, a

similar stretching as Figure 2 has occurred and the

output signals are all 46 frames. Yet the alignment

of the signals is regarded. The global maximum and

minimum points are aligned with higher accuracy as

they take place in the same time compared to Figure

2, and frames 20 to 30 have not been misaligned.

Comparing Figure 3 and Figure 4, we can observe

that Figure 4 shows more alignment in the case of

global minimum for frames 20 to 30.

Animating the new warped signals confirms the

fact that the conducted operations on the datasets

have not altered the data significantly and that the

primary or the secondary themes (k and r) are not

altered, thus the groundwork for computing the

conversion matrices has been laid successfully.

Figure 3: Signals from training, target, and test matrices

after time warping using maximum velocity features.

Figure 4: Signals from training, target, and test matrices

after time warping using minimum velocity features.

5 CONVERSION MATRIX

EXTRACTIONS AND

POST-PROCESSING

With the assumption that in (3), the k and r values

are unchanged and the same model is still valid, two

action sequences of the same class and different

secondary theme Y

[k,r

] and Y

[k,r

] can be

modelled by (8) and (9) respectively.

11111

][][][],[ erSrwkPrkY 





(8)

22222

][][][],[ erSrwkPrkY 





(9)

Differentiating among Y

and Y

generates ΔY

presented by (10).



 

121122

][][][][ eerSrwrSrwY 









(10)

MODELING AND TRANSFORMATION OF 3D HUMAN MOTION

311

Assuming equal weights for the secondary

themes and that only one secondary theme is present

in each sequence, the differentiated secondary

themes are the desired transformation matrix

between the two secondary themes, Γ[r

], which

is defined by (11) and computed by (12).

][][][][],[

112221

rSrwrSrwrr 

(11)



212221

],[],[],[ eerkYrkYrr 

(12)

In simple terms, for an action sequence with only

one secondary theme, the transformation matrix is

derived by differentiating the base and the target

sequences, along with noise elimination techniques.

It can be examined that the transformation

function T defined by (13) can be applied to action

sequence Y[k,r(1:f)] for converting the secondary

theme r

to r











jirr

rrwfrkYfrkYT

,]:1,[]]:1,[[



(13)

To eliminate the term (e

- e

) from (12), low

pass filtering is utilized. The affect of the noise

reduction process is presented in Figure 5.

The conversion signals are a result of

differentiating between two sequences of the same

primary theme with different secondary themes. Due

to the frame-to-frame approach for differentiation,

this process is very sensitive and sharp local and

global maximum and minimums are produced.

Animating the final raw outcome illustrates various

artefacts for different joints. Undesired motion such

as tremor and rigidity in some body parts are the

direct affect of the proposed techniques. Applying a

low pass filter results in smoothing of the conversion

data and the output signals as illustrated in Figure 5

eliminates some local maximum and minimums.

Disappearance of the tremor symptoms of the

animations as a result of filtering reaffirms the

necessity for post-processing the data.

Figure 5: The effect of low pass filtering.

Another technique was also employed for

transforming the secondary themes. With the

assumption that one secondary theme is present in

the sequence, the interpolation procedure derived by

(14) and (15) will also result in Y[k,r] which is a

sequence with the original primary action class P[k].





222111

],[],[

],[

rkYwrkYw

rkY











(14)





rSrwwrSrww

kPrkY 









222111

][][][][

][],[

(15)

In (15) the second term represents an

interpolation among the two secondary themes while

the third term e is the interpolated noise signal. Also

the first term P[k] indicates that the primary theme

of the action has remained unchanged. It is very

important that the secondary themes which are to be

interpolated be of the same nature. For instance, they

must be all related to age, or gender, or energy.

Interpolating between secondary themes of different

nature is meaningless as it is not logical to

interpolate, for instance, between a young theme and

a low energy theme. Nevertheless interpolation

between a young theme and an old theme is likely to

produce a mid-aged theme for the primary action

class k. The same low pass filtering process is

carried out to eliminate the noise from the output

data.

The two techniques presented above have been

implemented and the results are discussed and

compared in section 6. It is important to note the fact

that the conversion data do not need to include all

the joint values. For instance the markers placed on

the head don’t have a significant impact on the

secondary themes. A selected handful of markers

thought to be influential to the secondary themes are

altered and the rest are left unchanged, since

including all the markers will only increase the run

time and system error. The markers which have been

selected for manipulation are on the spine, arms, and

legs.

6 RESULTS AND DISCUSSIONS

As discussed in section 4 various techniques were

used for determining the feature instances for the

piece-wise time warping of the data. The minimum

velocity-based features proved to be most suitable

and are employed in the system. Successive to

warping the action sequences, the transformation of

GRAPP 2010 - International Conference on Computer Graphics Theory and Applications

312

the secondary themes take place using (8) to (15).

The computations are carried out in Matlab and the

results are transferred to Maya for visualization and

further evaluation.

Interpolation between the sequences was

implemented to compare with our approach. The

weights of 0.5 and 0.5 for w

and w

were selected

for each of the sequences to be interpolated. Figure 6

shows an original masculine jump (base) while

Figure 7 presents an original feminine jumping

(target) sequence successive to the time warping

procedure. Figure 8 illustrates the interpolated

output of the two actions. It is noticed that the action

is femininized to some extent, yet creating a 100%

feminine jump is only possible through setting the

weight of the masculine jump to zero. Although this

produces a perfect feminine jump, the dilemma is

the fact that the base action (masculine) is

completely excluded from the process; therefore the

system is only outputting the training feminine target

sequence. Figure 9 illustrates the output of the

system based on computing the transformation

function. In this sequence, the legs are femininized

similar to the interpolation output but with few

variations. The movement of the arms however, are

adjusted with more significance. Overall the output

is no worse, if not better, than the practical

interpolation. The huge advantage remains,

however, the fact that using the proposed model, we

have not eliminated the base action which the

secondary theme is desired to be added upon. The

same affect can be seen for other primary and

secondary themes and the advantage holds valid.

A rather detailed investigation for the proposed

transformation technique shows that the secondary

themes have been detached and added to the base

samples with high precision. Figure 10 (left) shows

the original masculine walk before conversion and

Figure 10 (right) shows the same keyframe after the

conversion. The Figure clearly shows the

transformation of feminine walk to masculine walk

using the model. While in feminine walk, the legs

are placed sequentially in front of one another, in

masculine walk, the legs are placed further apart.

Also the movement of the hip is limited in masculine

walk as opposed to the feminine where more

movement is visible in that area. Similar

manipulations are visible for other styles and other

classes of action such as that illustrated in Figure 11.

Figure 11 presents the conversion of low energy

(tired) run (left) to a high energy run (right) using

our proposed model. The movement of the legs

clearly shows the successful conversion of the two

styles.

Figure 6: Original masculine jump.

Figure 7: Original Feminine jump.

Figure 8: Interpolation output.

Figure 9: The output using transformation function.

To further evaluate the algorithms used to

synthesize the animation, a questionnaire was

formed and participants were asked to evaluate the

quality of the outputted animation. The

questionnaire included various enquiries regarding

each primary action class and all the three secondary

themes. The produced animation sequences were

highly acknowledged as only a few users were

unsatisfied with the results. An average 80%

approval rate by the participants regarding

successful style transformations using our model

confirms the significance of our proposed method.

Most of the concern with the animations were

regarding some of the running sequences which due

to the warping procedure, appeared as some sort of

fast-walk, yet re-runs of the animation and further

trials aided in clarifying and distinguishing between

the two classes. In general, our proposed method

was highly favoured over the interpolation

technique.

MODELING AND TRANSFORMATION OF 3D HUMAN MOTION

313

Figure 10: Original masculine walk (left), converted to

feminine walk (right).

Figure 11: Original low energy run (left), converted to

high energy run (right).

7 CONCLUSIONS AND FUTURE

WORK

The key insights into this paper are the piece-wise

time warping and features employed to conduct the

warping procedure, along with the mathematical

model proposed to describe human motion.

To align the motion samples correctly, critical

features were computed for each action sequence.

The signals were all warped such that after the

warping process they all contain the same number of

frames and the features be aligned i.e. they must

occur at the same time. The features were computed

in different fashions to search for the most

functional method. Selecting the feature instance

manually and using statistical analysis were the two

major categories where the second proved to be

much more efficient. Two different approaches were

taken based on statistical feature selection: the first

employed the instance when maximum net velocity

is reached, and the second when the minimum net

velocity occurs, where the second proved to be more

effective.

The proposed mathematical model is based on

dividing human motion into three components of

primary theme, secondary theme, and noise. The

primary theme is the main action class, for instance

walking, running or jumping and the secondary

theme is determined by the gender, energy, age, and

other specifications of the actor. The noise is

reduced by applying a low-pass filter.

Based on the model, a transformation function is

defined which can extract the secondary themes

from the training data and attach them to the test

sequence. The computed transformation signals are

used to convert any action of the same primary

theme to an action containing the desired secondary

theme.

For future work, further primary and secondary

themes are to be tested. Actions such as idling,

strafing left and right, crouching, and a collapsing or

death sequence which are popular in the field of

animation and game design are projected to be

included. Some other secondary themes such as

weight, happy/sad, and kind/angry also seem

important and worth further attention.

Other transformation functions based on the

proposed model could be derived and applied. As

the model seems promising, further study on

developing new techniques for separation of the

themes is necessary.

Another area in which this research seems

promising is the combination of secondary themes.

It is usually necessary to produce combined styles

such as old-angry, or young-feminine and etc. We

are currently working on different methods for

combining the transformations for each style such as

weighted averaging, MIN/MAX, and rule-based

operators.

REFERENCES

Brand, M., Hertzmann, A., 2000. Style Machines. In

Proceedings of the 27th International Annual

Conference on Computer Graphics and Interactive

Techniques. July 2000, pp. 183-192.

Egges, A., Molet, T., Magnenat-Thalmann, N., 2004.

Personalised Real-time Idle Motion Synthesis. In

Proceedings of the 12th Pacific Conference on

Computer Graphics and Applications. Oct 2004, pp.

121-130.

Fang, A. C., Pollard, N. S., 2003. Efficient Synthesis of

Physically Valid Human Motion. ACM Transactions

on Graphics. 22(3), pp. 417-426.

Hsu, E., da Silva, M., Popovic, J., 2007. Guided time

warping for motion editing. In Proceedings of the

2007 ACM SIGGRAPH/Eurographics symposium on

Computer animation, August 2007, pp. 45-52.

Hsu, E., Pulli, K., Popovic, J., 2005. Style Translation for

Human Motion. In ACM SIGGRAPH ’05. July 2005,

pp. 1082-1089.

GRAPP 2010 - International Conference on Computer Graphics Theory and Applications

314

Hutchinson, A., 1996. Labanotation, Dance Books.

Kang, H., Kim, M., 2007. A Study of Walking Motion for

Game Character with a Player’s Emotional Factors

Applied. In 5th ACIS International Conference on

Software Engineering Research, Management &

Applications, Aug 2007, pp. 911-916.

Li, Y., Wang, T., Shum, H. Y., 2002. Motion Texture: A

Two-Level Statistical Model for Character Motion

Synthesis. ACM Transactions on Graphics. 21(3), pp.

465-472.

Liu, C. K., Popovic, Z., 2002. Synthesis of Complex

Dynamic Character Motion from Simple Animations.

In Proceedings of the 29th International Annual

Conference on Computer Graphics and Interactive

Techniques. July 2002, pp. 408-416.

Pullen, K., Bregler, C., 2002. Motion Capture Assisted

Animation: Texturing and Synthesis. ACM

Transactions on Graphics. 21(3), pp. 501-508.

Rose, C., Cohen, M. F., Bodenheimer, B., 1998. Verbs and

Adverbs: Multidimensional Motion Interpolation.

IEEE Computer Graphics and Applications. 18(5), pp.

32-40.

Safonova, A., Hodgins, J. K., Pollard, N. S., 2004.

Synthesizing Physically Realistic Human Motion in

Low-Dimensional, Behavior-Specific Spaces. ACM

Transactions on Graphics. 23(3), pp. 514-521.

Sok, K., Kim, M., Lee, J., 2007. Simulating Biped

Behaviors from Human Motion Data. ACM

Transactions on Graphics. 26(3), Article 107.

Su, W., Pham, B., Wardhani, A., 2005. High-level Control

Posture of Story Characters Based on Personality and

Emotion. In Proceedings of the 2nd Australasian

Conference on Interactive Entertainment, Nov 2005,

pp. 179-186.

Tanco, L. M., Hilton, A., 2000. Realistic synthesis of

novel human movements from a database of motion

capture examples. In Proceedings of Workshop on

Human Motion, July 2000, pp. 137-142.

MODELING AND TRANSFORMATION OF 3D HUMAN MOTION

315