Risk-Oriented and Explainable Framework for Supporting Judging in Combat Sports: Preliminary Assessment of Striking and Non-Striking Movement Detection

Risk-Oriented and Explainable Framework for Supporting Judging in Combat Sports: Preliminary Assessment of Striking and Non-Striking Movement Detection

leila nasiri darani1 amir foad ebrahimi por2 hadiye hamidi mishamandani3 samane mohammadi4 shilan mahmodiyan5

1) Iran, Guilan, Lahijan, Islamic Azad University, of Artificial Intelligence
2) Iran, Guilan, Lahijan, Islamic Azad University, of Artificial
3) Iran, Guilan, Lahijan, Islamic Azad University, of Artificial
4) Iran, Guilan, Lahijan, Islamic Azad University, of Artificial
5) Iran, Guilan, Lahijan, Islamic Azad University, of Artificial

IOI Code : AAHD-ABKDZH
Publication : National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
    Judging in combat sports is challenging due to the high speed of movements, occlusion, changes in viewing angles, and the difficulty of reliably determining physical contact. These challenges create a need for decision-support systems that provide reviewability and reliability. This study aims to propose and evaluate a risk-oriented and explainable framework for supporting judging decisions based on binary detection of striking and non-striking movements in combat sports clips. The dataset used in this study consists of 30 balanced clips that were reviewed by individuals with specialized experience in combat sports. Within the proposed framework, four feature-extraction streams were investigated: HOG-SVM as a classical baseline, SlowFast video features, pose-sequence features, and a hybrid Fusion model. To provide a more robust evaluation of model performance, five-fold cross-validation was employed. In addition to conventional classification metrics, probability calibration, high-risk errors, selective classification, and the referral of uncertain cases to a human judge were also examined. Furthermore, exploratory analyses related to joint movements, the temporal dynamics of skeletal data, and outliers were conducted to investigate the quality and interpretability of the extracted features. The results showed that the observed superiority of the Fusion model in the single-split evaluation was not confirmed after cross-validation, while Pose-Sequence demonstrated the best relative performance on this small dataset. This finding indicates that the evaluation of intelligent judging-support systems should not rely solely on classification accuracy or the results obtained from a single data split. In such applications, the stability of results, probability calibration, identification of high-risk errors, quality control of skeletal data, and referral of uncertain cases to human judges should also be taken into consideration. Therefore, the aim of the present study is not to introduce a system that is ready for fully automated judging, but rather to provide preliminary evidence supporting the necessity of adopting a risk-oriented, explainable, and human-in-the-loop approach when designing intelligent decision-support systems for sports judging. Given the limited size of the dataset, the calibration, risk-analysis, and explainability findings reported in this study should be considered preliminary and exploratory and should therefore be interpreted with caution.

Keywords : intelligent judging support combat sports movement detection risk analysis cross validation calibration human in the loop