Page 41 - 2024F

34 UEC Int’l Mini-Conference No.53

based on body mechanics. However, challenges
remain in distinguishing nuanced pitch types, such
as changeups and curveballs, due to their subtle and
deceptive body movements.
In binary classification tasks, the method showed
robustness in distinguishing fastballs from non-
fastballs, with an accuracy of 85.7%, highlighting
its effectiveness in detecting pitches with straight-
forward mechanics. Nevertheless, distinguishing
fast and slow pitches proved more challenging,
(a) achieving an accuracy of 80.8%. Misclassifications
often occurred with changeups, which exhibit ex-
plosive initial mechanics similar to fastballs before
decelerating, complicating the classification pro-
cess. These results underscore the complexity of
fine-grained pitch type recognition and emphasize
the need for further refinement in handling decep-
tive motion patterns.

V. CONCLUSIONS
(b) This work demonstrates the feasibility and effec-
tiveness of a skeleton-based approach for baseball
pitch classification, achieving competitive accuracy
without relying on ball trajectory data. The focus
on body mechanics provides a novel perspective,
particularly valuable in scenarios where specialized
tracking systems are unavailable.
While the results are promising, particularly in
binary classification tasks, further advancements
are necessary to address the challenges posed by
nuanced and deceptive pitch types. Future direc-
(c) tions include exploring additional biomechanical
features, such as joint velocities and limb angles,
Fig. 2: Confusion matrices for (a) Six-Class Pitch to enhance motion representation. Additionally, in-
Type Classification, (b) Fastball vs. Non-Fastball tegrating ball trajectory data through a two-stream
Classification, and (c) Fast vs. Slow Pitch Classifi- framework could provide complementary insights,
cation. Note: The values are expressed in percent- potentially improving classification performance.
ages.
By advancing techniques for analyzing human
movement in baseball, this study contributes to
both sports analytics and the broader field of fine-
for the use of broadcast footage, making it more grained action recognition.
accessible for practical applications.
Our findings reveal the potential of this ap-
proach, particularly in isolating the pitcher’s move- VI. ACKNOWLEDGMENTS
ments, thereby minimizing noise from background
elements or other individuals in the frame. The The author expresses gratitude to the Japanese
six-class classification task achieved a promising and Mexican governments for their support through
overall accuracy of 68.2%, demonstrating the capa- the JASSO and CONAHCYT scholarships, respec-
bility of the model to differentiate between pitches tively.