On Automatic Actions Retrieval Of Martial Arts

2004 IEEE International Conference on Multimedia and Expo (ICME)

On Automatic Actions Retrieval of Martial Arts

Timothy K. Shih', Ching-Sheng Wang', Yuan-Kai Chiul, Yi-Tsou Hsin', and Chun-Hong Huang'

'Department of Computer Science and Information Engineering

Tamkang University, Taiwan, R.O.C.

tshih@cs.tku.edu.tw

2Department of Computer and Information Science

Aletheia University, Taiwan, R.O.C.

cswanetdemail .au.edu.hv

ABSTRACT

Martial art actions can he represented via VRML

animations or extracted by video tracking. We propose an

action retrieval method, which allows users to retrieve

similar actions of martial arts. The mechanism is based on

a similarity function that compares animation uacks. A

representation of human skeleton includes head, knee,

elbow, wrist, etc further aggregates important features in

martial art actions. Different weights are dynamically

calculated according to motion sensitivity of feature points.

As a result, the system can automatically retrieve similar

martial art actions. The results are tested by professional

kung fu master with a good satisfaction.

Key words: VRML, animation, automatic action

retrieval, virtual reality, martial art

1. INTRODUCTION

Automatic retrieval of actions in 3D space is a challenge

but useful technique. Examples of behavior understanding

of video can be found in [I, 21. Motion tracking and

recognition of human interactions by a multi-layer finite

state machine is presented in [l]. By using body pose

vectors, human action recognition is presented in [2].

With recent Virtual Reality technologies, movies can he

made my VR-based or Augmented Reality-based actors.

Retrieval of actors in a 3D scene become a useful

technology [3,5], if animated actors in a scene datahase is

to be reused [4]. In stead of using user pre-defined

metadata [4], 3D animations should be retrieved based on

the existing animation models. We look at one particular

domain as an example showing our contribution and its

possible extensions. Martial arts can be represented as

VRML animations. A particular martial art, known as the

Bar-Chi Spar, is our target application. Bar-Chi Spar was

originated in Ho-Bai, China, in around 1368. The unique

basic action of Bar-Chi Spar is clear, fast, powerful, and

smooth. According to our Bar-Chi Spar guru, the use of

head, shoulder, elbow, pud, tail, crotch, knee, and foot are

the features which represent different sets of spar actions.

Assuming that these features can be extmcted from a

VRML model, which teaches Bar-Chi Spar, it is possible

to record the animation tracks of these important portions

of a human body and tn save it for comparison. Thus, an

automatic retrieval system tells the user which set of spar

actions is similar to the one heishe is learning.

In order to compare the animation tracks of different

feature points, a normalization technique is required since

different VRML actors may have different highs and

weights. Also, animation tracks may have different

lengths. We present the normalization of action tracks in

section 2. A skeleton is required to represent human body,

which is presented in section 3. A distance function which

aggregates feature points in Bar-Chi Spar is presented in

section 4. Our system is implemented using the 3D Studio

Max and the Cortona VR Player. We compare the

retrieval outcome with the reviews from three kung fu

masters in section 5, before our conclusion section is

presented.

2. NORMALIZATION OF ACTION TRACKS

Object animation tracks are not necessary represented

with the same amount of tracking points. It is necessary to

normalize animation tracks before we use these tracks in a

similarity function. One way to normalize number of

tracking points in two tracks is to add interpolation points.

According to 3D geometric, three feature points form a

circle in the three dimensional space. With a limited

granularity, it is reasonable to use interpolated points on

the circle as an approximation of an animation track.

Assuming that we have three points Pl=(xl,yl,zl),

P2=(x2,y2,z2), and P3=(x3,y3,z3). Also, let point (X, Y, 2)

...