Research Proposal: Expanding Psychomotor

Performance Modeling in Karate Using Computer

Abstract
This research aims to extend the existing work on modeling psy-
chomotor performance in Karate using computer vision pose estimation.
Building upon the study ”Toward Modeling Psychomotor Performance in
Karate Combats Using Computer Vision Pose Estimation,” which focused
on Ippon Kihon kumite, this proposal seeks to analyze Sambon Kihon ku-
mite, a more complex three-step conventional assault. By leveraging the
OpenPose algorithm for feature extraction and applying machine learning
and deep learning techniques for classification, this research will expand
the dataset and improve the accuracy of movement recognition. The study
will compare various classification algorithms, assess their computational
efficiency, and introduce sub-activity recognition to model movement se-
quences. The expected outcomes include a more comprehensive dataset,
enhanced classification accuracy, and deeper insights into the application
of AI-driven psychomotor learning in martial arts.

1 Introduction
This research proposal outlines a project that builds upon the work presented
in the study ”Toward Modeling Psychomotor Performance in Karate Combats
Using Computer Vision Pose Estimation” (Echeverria & Santos, 2021). The
original study explored the use of computer vision techniques, specifically the
OpenPose algorithm, to identify postures within a Karate kumite, focusing on
Ippon Kihon kumite, a one-step conventional assault. This proposal suggests
extending this approach to Sambon Kihon kumite, a three-step conventional
assault, to create a more comprehensive dataset and to further analyze the
performance of different machine learning and deep learning algorithms. This
will provide a deeper understanding of psychomotor performance in martial arts
and the capabilities of computer vision for this purpose.

1
2 Research Gap
Existing research in human movement computing, including that focused on
martial arts, largely concentrates on individual movements. There is a lack
of studies that model the movements performed during the joint practice of
multiple users, where variables such as orientation, fatigue, adaptation to a
partner’s anatomy, and affective state can be relevant. Furthermore, there is a
gap in research modeling motion to evaluate psychomotor performance in a way
that provides appropriate feedback for technique improvement. Current systems
also do not address the problem of modeling movements from a psychomotor
perspective, nor do they compare a user’s movement with their previous perfor-
mance or with the performance of users of different expertise levels. The initial
research demonstrated the use of OpenPose to extract key points from videos
of karate practitioners, which were then used to train data mining algorithms
to classify and identify basic karate movements. However, the study focused on
static postures and basic Ippon Kihon kumite movements. The dataset used
was limited in scope, concentrating on single-step movements and lacking an
analysis of movement sequences.

3 Background
Human activity recognition (HAR) techniques focus on recognizing and classi-
fying inputs using sensory signals, images, or video to determine the type of
activity a person is performing. These techniques are usually divided into mod-
els based on image or video and those based on signals collected from sensors
such as accelerometers and gyroscopes. Models based on images and videos
provide information about the movements performed by the skeletal structure
of the body. The martial arts domain is a useful context for research on psy-
chomotor activity modeling, as it requires the development and refinement of
psychomotor skills for progress in practice. Some studies have applied move-
ment modeling to analyze techniques performed by karate practitioners, but
these have focused solely on individual movements. The initial research utilized
the OpenPose algorithm for pose estimation due to its open-source nature, prac-
tical applicability, and large developer community. This research successfully
demonstrated the use of OpenPose to extract key points from videos of karate
practitioners, which were then used to train data mining algorithms to classify
and identify basic karate movements. The study focused on static postures and
basic Ippon Kihon kumite movements, such as Kamae, Gedan Barai, Oi Tsuki,
Soto Uke, and Gyaku Tsuki. The results indicated high classification accuracy
using algorithms such as BayesNet, J48, MLP, and DeepLearning4J (Echeverria
& Santos, 2021).

2
4 Problem Definition
Current research on human movement analysis, particularly in martial arts, is
limited by several factors. There is a lack of datasets that capture complex,
multi-step movements in a martial arts context. Existing models do not assess
psychomotor performance over time or in comparison with other practitioners.
Additionally, few studies address the modeling of movements during the joint
practice of two individuals, which introduces complexities related to interaction
and adaptation. The scope of existing datasets is limited, as they focus on
individual movements and do not model movement sequences.

5 Objectives
This research aims to address these limitations by expanding the dataset of
karate movements to include Sambon Kihon kumite, thereby providing a more
robust test for classification algorithms. The study will evaluate the accuracy
of computer vision techniques in identifying postures and movement sequences
within Sambon Kihon kumite. Additionally, it will compare the performance
and computational costs of different machine learning algorithms (BayesNet and
J48) and deep learning algorithms (MLP and DeepLearning4J) when analyzing
this more complex dataset. The research will enhance the system’s capacity to
identify sequences of movements through sub-activity recognition within Sam-
bon Kihon kumite. Finally, the results of this study will be compared with
those obtained from the Ippon Kihon kumite dataset to assess the impact of
increased data complexity on algorithm performance.

6 Methodology
The research methodology consists of several key steps. First, data acquisition
will involve capturing video recordings of karate practitioners performing Sam-
bon Kihon kumite using the KUMITRON system, which integrates video and
sensor data to capture movement. Both static and dynamic recordings will be
collected from different angles to capture a wider range of movements. Data will
be collected from multiple karate practitioners with varying levels of experience.
Feature extraction will be conducted using the OpenPose algorithm, which
will extract key point data representing the 2D coordinates of body joints. These
key points will be used to generate 26 triplets corresponding to the main anatom-
ical parts and to calculate angles between body segments.

7 Expected Outcomes
This research is expected to yield several key outcomes. It will produce an ex-
panded dataset of karate movements, including Sambon Kihon kumite, which

3
will be made available for public use. Additionally, it will provide a compre-
hensive performance analysis of machine learning and deep learning algorithms
in the context of complex karate movement recognition. The study will evalu-
ate the generalizability of the existing system to accommodate more complex
martial arts movements. Moreover, it will contribute to a deeper understanding
of the potential of computer vision and AI for psychomotor learning in martial
arts. Finally, the research will develop a system capable of identifying move-
ment sequences (sub-activity recognition) within Sambon Kihon kumite, a novel
contribution to the field.

8 References
Echeverria, J., & Santos, O. C. (2021). Toward Modeling Psychomotor Perfor-
mance in Karate Combats Using Computer Vision Pose Estimation.