Author(s): CHEN, H.H., TSAI, C.Y., Institution: TAITUNG UNIVERSITY, Country: TAIWAN, Abstract-ID: 532

INTRODUCTION:
Effective gripping techniques play a crucial role in success of judo, serving as the foundation for initiating throws or attacks. In a very limit number of studies exploring the gripping techniques preceding scored throws.This study aimed to examine the game performance of female athletes in the category of 52kg, specifically focusing on their gripping behavior before executing throwing techniques.Additionally, a Multi-layer Perceptrons (MLP) model and a CycleGAN are constructed to distinguish between effective and ineffective gripping techniques
METHODS:
Analysis was conducted on 100 videos of female athletes ranked in the top 50, taken from the world circuit tournament spanning 2019-2022. The process were notated by 5 experienced observes.
RESULTS:
(1) Time analysis revealed a mean combat time of 261.5 secs (±128 secs) with an average of 9.1 (±5.2) attacks. Attack duration decreased over sections, from 18.5(±7.3secs) to 14.4 secs. The effort-pause (EP ratio) also reduced from 4 to 2.5 (2). Motions analysis demonstrated that 91% of gripping before throwing techniques did not show obvious effectiveness out of 593 attacks. The afterward throwing techniques executed almost same side as the gripping force motion particular for the foot and hip skills (97%). However, only 32.2 % of throwing techniques got the advantage consequence and over half of them without advantage (66%). Although the foot and hand claimed the majority of the percentage of skills (58% and 22%, respectively), the hip and sacrifice skills had a lager efficiency ratio (numbers of throws / number of advantage) for gaining the attack advantage (1.3 and 0.74). (3) Multi-state Markov model identified primary gripping behaviors preceding the throwing motions, such as grasping approaching sleeves or grasping the right collar then left sleeve. (4) The initially-constructed MLP gripping model achieved a mean test accuracy of 67.72±0.93 and a mean test F1-score 79.09±0.46 in discriminating the attack consequences. As we pre-train another CycleGAN model to generate adequate augmented data and use it to train the MLP along with the original data, we get an improved mean test accuracy and a mean F1-score up to 73.68±0.00 and 81.29±0.013, respectively.
CONCLUSION:
(1) The decline of EP ratio suggested more time for attempting and strategic planning at the beginning of combat, leading a stable gripping behavior pattern. (2) Almost none effective gripping preceding the initiation of attacks may implicate the brinkmanship behaviors applied by the athletes in the interpersonal synergy of judo combat. (3) Limited combat data sets may affect the accuracy of the gripping model. However, the data generation mechanism could be an alternative that can produce enough and effective augmented data to facilitate the classification model training for better generalization. Therefore, the intricate combination of griping and throwing motion emerges as a critical skills for elite judo athletes. An accurate gripping model could be built.