TY - GEN
T1 - Exploring machine learning techniques for penalty kick image classification
T2 - 5th International Conference on Information Technology and Contemporary Sports, TCS 2025
AU - Su, Guanpeng
AU - Liang, Zijun
AU - Ho Tat, U.
AU - Wu, Kairi
AU - He, Ziyue
AU - Wong, Dennis
N1 - Publisher Copyright:
© COPYRIGHT SPIE. Downloading of the abstract is permitted for personal use only.
PY - 2026/2/12
Y1 - 2026/2/12
N2 - This study examines the effectiveness of classical machine learning models in classifying penalty kick images extracted from video footage. The training and primary testing were based on frames captured from YouTube videos of penalty kicks. To evaluate the models' generalization ability, a separate test set was introduced, consisting of frames taken from self-recorded penalty kick videos filmed under real-world conditions. Seven widely used classifiers - Logistic Regression, Naive Bayes, Random Forest, Decision Tree, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Linear Discriminant Analysis (LDA) - were assessed using consistent preprocessing procedures. Among these, the Random Forest model delivered the best performance on both datasets, achieving an accuracy of 67.75% on the YouTube-based set and 63.33% on the self-filmed set. These findings demonstrate the robustness and cross-domain adaptability of ensemble learning methods, particularly Random Forest, in sports image classification tasks. The results provide a practical reference for developing lightweight and interpretable models in scenarios where deep learning may be less feasible.
AB - This study examines the effectiveness of classical machine learning models in classifying penalty kick images extracted from video footage. The training and primary testing were based on frames captured from YouTube videos of penalty kicks. To evaluate the models' generalization ability, a separate test set was introduced, consisting of frames taken from self-recorded penalty kick videos filmed under real-world conditions. Seven widely used classifiers - Logistic Regression, Naive Bayes, Random Forest, Decision Tree, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Linear Discriminant Analysis (LDA) - were assessed using consistent preprocessing procedures. Among these, the Random Forest model delivered the best performance on both datasets, achieving an accuracy of 67.75% on the YouTube-based set and 63.33% on the self-filmed set. These findings demonstrate the robustness and cross-domain adaptability of ensemble learning methods, particularly Random Forest, in sports image classification tasks. The results provide a practical reference for developing lightweight and interpretable models in scenarios where deep learning may be less feasible.
KW - Cross-Domain Classification
KW - Machine Learning
KW - Penalty Kick Classification
KW - Random Forest
KW - Sports Video Analysis
UR - https://www.scopus.com/pages/publications/105032919483
U2 - 10.1117/12.3103395
DO - 10.1117/12.3103395
M3 - Conference contribution
AN - SCOPUS:105032919483
T3 - Proceedings of SPIE - The International Society for Optical Engineering
BT - Fifth International Conference on Information Technology and Contemporary Sports, TCS 2025
A2 - Yaakob, Razali
A2 - Akilan, Thangarajah
PB - SPIE
Y2 - 14 November 2025 through 16 November 2025
ER -