Skill Identification Using Time Series Data Mining

Document technical information

Format pdf
Size 180.7 kB
First found May 22, 2018

Document content analysis

Category Also themed
Language
English
Type
not defined
Concepts
no text concepts found

Persons

Organizations

Places

Transcript

68
Int'l Conf. Data Mining | DMIN'16 |
Skill Identification Using Time Series Data Mining
Toshiyuki MAEDA1 and Masumi YAJIMA2
1
Faculty of Management Information, Hannan University, Japan
2 Faculty of Economics, Meikai University, Japan
Keywords:
Time Series Data, Sports Skill, Data Mining, Motion Picture, Knowledge Acquisition
Abstract:
This paper addresses sports skill identification using time series motion picture data, focused on volleyball. In
this paper, volleyball play is analyzed with motion picture data recorded by hi-speed cam-coder, where we do
not use physical information such as body skeleton model, and so on. Time series data are obtained from the
motion picture data with marking points, and analyzed using data mining methods such as Naive Bayes, and
other tree learning algorithm. We attempt to identify technical skill models of volleyball attacks.
1
Introduction
2 Related Works
For not only engineering skills but also sports
skills, many researchers treat body structure models
and/or skeleton structure models obtained from physical information such as activity or biomechatronical
data for sports skill researches [1].
Those might be because those researchers believe
that internal models of technical skill are structured
physically, with some skill levels which are human
intention, environmental adjustment, and so on [2].
For instance, Matsumoto and others describes
skilled workers skills which have actually internal models of structured skill architecture and they
choose an action process from internal models adjusted with environment [3]. It is even though hard
for skilled workers to represent internal models by
themselves. They reflect involuntarily their own represented actions, and achieve highly technical skills
with internal models.
We had, however, researched that fore-hand
strokes of table tennis play exemplify sports action,
and classify skill models using motion picture data
analysis without body structure model nor skeleton
structure model. We had evaluated those into three
play levels as expert/intermediate/novice, and classify the models using data mining technologies [4, 5].
That means this research is a challenge to clarify internal models only from represented image data.
We hence have an attempt to apply our research
framework to other sports skills, and then this paper
addresses a personal sports skill identification using
time series motion picture data, focused on volleyball.
Wilkinson[6] describes that qualitative skill analysis is an essential analytic tool for physical educators
and refers to a process in which a teacher identifies
discrepancies between the actual response observed
and the desired response. Providing instruction for
preserving teachers regarding how to recognize errors has been largely neglected in teacher preparation.
The purpose of this study was to evaluate an alternative approach for teaching qualitative skill analysis
to undergraduates. The study evaluated the effectiveness of a visual-discrimination training program. The
subjects were 18 undergraduate students. The visualdiscrimination training program was introduced using a multiple-baseline design across three volleyball skills: the forearm pass, the overhead pass, and
the overhead serve. After the introduction of each
instructional component, subjects made abrupt improvements in correctly analyzing the volleyball skill.
This approach for teaching qualitative skill analysis
is one alternative to the conventional techniques currently being used in professional preparation.
Watanabe et al.[7] shows a method for the measurement of sports form. The data obtained can be
used for quantitative sports-skill evaluation. Here,
they focus on the golf-driver-swing form, which is
difficult to measure and also difficult to improve. The
measurement method presented was derived by kinematic human-body model analysis. The system was
developed using three-dimensional (3-D) rate gyro
sensors set of positions on the body that express the
3-D rotations and translations during the golf swing.
The system accurately measures the golf-driver-swing
form of golfers. Data obtained by this system can be
ISBN: 1-60132-431-6, CSREA Press ©
Int'l Conf. Data Mining | DMIN'16 |
69
related quantitatively to skill criteria as expressed in
respected golf lesson textbooks. Quantitative data for
criteria geared toward a novice golfer and a mid-level
player are equally useful.
Barzouka et al.[8] examine the effect of feedback
with simultaneous skilled model observation and selfmodeling on volleyball skill acquisition. 53 pupils
12 to 15 years old formed two experimental groups
and one control group who followed an intervention
program with 12 practice sessions for acquisition and
retention of how to receive a ball. Groups received
different types of feedback before and in the middle
of each practice session. Reception performance outcome (score) and technique in every group were assessed before and at the end of the intervention program and during the retention phase. A 3 (Group) ×
3 (Measurement Period) multivariate analysis of variance with repeated measures was applied to investigate differences. Results showed equivalent improvement in all three groups at the end of the intervention
program. In conclusion, types of augmented feedback
from the physical education teacher are effective in
acquisition and retention of the skill for reception in
volleyball.
3
Experiments
Our research is to identify internal models from
observed motion picture data and skill evaluation with
represented actions, without measurement of the body
structure or the skeleton structure.
We focus on volleyball attack among various
sports, and analyze volleyball skills of stacks from
observed motion picture data and skill evaluation with
represented actions.
For the feasibility research, we have recorded motion pictures of 6 subjects who are 3 expert / 3 novicelevel university students. As skill evaluation of representing action, We classify the levels as follows;
and frame-rate: 300 frames per seconds) installed besides of the players. On playing in 5 minutes, several
attack motions are recorded for each player.
4 Skill Identification
From the recorded motion pictures, 100 to 200
frames are retrieved from the beginning of take-back
to the ball until the end of the attack. We have then
distributed two dimensional axes positions (pixel values) of 4 marking points for each frame, where the
starting point is set at the shoulder position of the first
frame.
We then attempt an investigation using data mining technique. The skill evaluation of representing
action consists of two classes such as Expert and
Novice. Each marking position is represented two dimensional and so the observed data are reconstructed
in 48-input / 2-class output.
For applying observed data of fore-hand strokes of
6 subject players, we reconstruct time series data from
the original data. One datum is a set of 48-tuple numbers such as 4 markings × 2 axis (x, y)×6 frames, and
each datum is overlapped with 3 frames data (from
fourth to sixth frame) of the next datum for presenting linkage of each datum (see Figure 2).
We use an integrated data mining environment
“weka” [9] and analyze the data by analyzing methods of J48 (an implementation of C4.5) and NBT
(Naive Bayes Tree).
4.1 Pre-examination
We have had a pre-examination for applying our research framework into volleyball skills. Table 1
shows the recognition rate of the data sets. As for
expert players, two third of data on three players are
used as learning data, and the rest for evaluation.
Table 1: Recognition rate of modified data sets.
• Expert class: members of volleyball club at university,
• Novice class: inexperienced students.
Figure 1 shows positions of marking setting. Each
player is marked at 4 points on the right arm as;
1. Left knee,
2. Right waist,
3. Right shoulder, and
4. Right elbow.
We have recorded swing traces of stacks using a
high-speed cam-corder (resolution: 512 × 384 pixel
J48
NBT
Recognition Rate(%)
Cross validation
Learning and test
95.9
48.2
97.7
62.7
Table 2 shows the discrimination of classes of
NBT classification.
In those results, the recognition rates for evaluation data are not so good, though NBT makes better
results for evaluation data. On the contrary, the result
of the number of class recognition for each method in
Table 2 implies that NBT tend to recognize Expert as
Novice. This implies that, even in Expert class, data
ISBN: 1-60132-431-6, CSREA Press ©
70
Int'l Conf. Data Mining | DMIN'16 |
FIGURE 1: Measurement markings.
Table 2: Discrimination of classes for NBT classification.
Expert
Novice
Classified as
Expert
Novice
66
58
14
55
may have some variation. We then focus on personal
identification, as data variation may cause from personal skill variation e.g. an expert position in plays
such as attacker, center player, and so on.
4.2
Personal Identification
As mentioned above, we here analyze for personal
identification, where two subjects are compared for
each test. One test is with an Expert and a Novice,
and the other test is with two Experts. We here use
NBT for these tests as NBT are better than J48 on
above tests.
These results are fairly good and suggest this anal-
Table 3: Recognition rate of personal data sets.
One Expert and
one Novice
One Expert and
another Expert
Recognition Rate(%)
Cross
Learning
validation and test
96.9
97.5
94.2
100.0
ysis process can identify each person, especially identification for two Experts, though we need further investigation.
5 Conclusion
This paper addresses analysis and classification
for internal models for technical skills as evaluation
skillfulness for volleyball attack motion, and discuss
skill identification. We had some experiments and
ISBN: 1-60132-431-6, CSREA Press ©
Int'l Conf. Data Mining | DMIN'16 |
71
FIGURE 2: Data structure from isolated pictures.
some results imply that expert or intermediate players can make some categorical groups for technical
skills, but there seems not to be a category for novice
players because of various individual technical skills.
Furthermore, for applying observed data of volleyball stacks of players, we reconstruct time series data
from the original data and analyze the new data by
data mining techniques such as J48, NBT, where the
recognition rate for evaluation data is fairly good, and
NBT makes better results for learning and evaluation
data. Personal analysis furthermore may be better categorized. As future plans, we have to progress further
experiments, and measure more precise data and then
analyze if needed.
Acknowledgment
Part of this research was supported by JSPS KAKENHI Grant Number 15K02185 and 15K03802. This
research was assisted, especially for data management, by Mr. Y. Tamari and Mr. Y. Tsujino at Hannan
University. The authors greatly appreciate those.
References
[1] Y. Mochizuki, R. Himeno, and K. Omura: Artificial Skill and a New Principle in Sports (Special Issue on Digital Human : Measurement and
Modeling of Human Functions). (in Japanese)
System, Control, and Information, Vol.46, No.8,
pp.498–505 (2002)
[2] T. Shiose, T. Sawaragi, K. Kawakami, and O.
Katai: Technological Scheme for the Preservation of Technique from Ecological Psychological Approach. (in Japanese), Ecological Psychology Research, Vol.1, No.1, pp.11–18 (2004)
[3] Y. Matsumoto: Organization and Skill – Organization Theory of Preservation of Technique (in
Japanese), Hakuto Shobo (2003)
[4] T. Maeda, I. Hayashi, and M. Fujii: Time Series
Data Analysis for Sport Skill. Proceedings of the
2012 12th International Conference on Intelligent Systems Design and Applications , pp.392–
397, Kochi, India (2012)
[5] T. Maeda, I. Hayashi, M. Fujii and T. Tasaka:
Sport Skill Classification Using Time Series
Motion Picture Data. Proceedings of thethe 40th
Annual Conference of the IEEE Industrial Electronics Society (IECON 2014) , pp.5272–5277,
Dallas, TX USA (2014)
ISBN: 1-60132-431-6, CSREA Press ©
72
Int'l Conf. Data Mining | DMIN'16 |
[6] S. Wilkinson: A Training Program for Improving Undergraduates’ Analytic Skill in Volleyball. Journal of Teaching in Physical Education,
Vol.11, Iss.2, pp.177–194 (1992)
[7] K. Watanabe, and M. Hokari: Kinematical analysis and measurement of sports form. IEEE
Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans, Vol.36, Iss.3,
pp.549–557 (2006)
[8] K. Barzouka, N. Bergeles, and D. Hatziharitos: EFFECT OF SIMULTANEOUS MODEL
OBSERVATION AND SELF-MODELING OF
VOLLEYBALL SKILL ACQUISITION. Perceptual and Motor Skills, Vol.104, pp.32–42
(2007)
[9] http://www.cs.waikato.ac.nz/ml/weka/ (2014)
ISBN: 1-60132-431-6, CSREA Press ©
×

Report this document