Deteksi Pengenalan Aksi Sepak Bola dengan Pendekatan Deep Learning

Adinata, Ivan Fairuz (2026) Deteksi Pengenalan Aksi Sepak Bola dengan Pendekatan Deep Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221167-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy

Abstract

Deteksi aksi pada video sepakbola menghadirkan tantangan yang kompleks, meliputi pergerakan yang cepat, oklusi objek, dan adanya ketidakseimbangan distribusi kelas pada dataset. Untuk dapat mengatasi permasalahan tersebut, diperlukan model deep learning dan metode yang dapat mengenali dan membedakan setiap kelas secara baik. Metode yang diusulkan mengintegrasikan arsitektur Temporal-Discriminability Enhancer Encoder-Decoder (T-DEED) sebagai baseline yang diperkuat dengan Adaptive Spatio-Temporal Refinement Module (ASTRM) pada backbone RegNet-Y. Selain itu, fungsi Soft Instance Contrastive (SoftIC) Loss juga diterapkan secara khusus untuk menangani masalah class imbalance. Metodologi penelitian merancang lima skenario pengujian untuk menguji efektivitas integrasi ASTRM dan SoftIC Loss dibandingkan dengan model baseline. Kinerja model diukur menggunakan metrik mean Average Precision (mAP) dengan toleransi temporal 1 detik sesuai protokol evaluasi SoccerNet. Hasil pengujian pada skenario terbaik berhasil menaikkan performa mAP menjadi 57,75% dari performa baseline T-DEED sebesar 53,32% mAP. Hasil penelitian diharapkan memberikan kontribusi dalam peningkatan akurasi deteksi aksi sepak bola, khususnya pada kondisi class imbalance, serta memperkaya kajian action spotting berbasis video olahraga, khususnya video sepakbola.
========================================================================================================================================
Action Spotting in football videos presents complex challenges, including fast movements, object occlusion, and class imbalances in the dataset. To address these issues, deep learning models and methods capable of accurately recognizing and distinguish each class are required. The proposed method integrates the Temporal-Discriminability Enhancer Encoder-Decoder (T-DEED) architecture as a baseline, reinforced with the Adaptive Spatio-Temporal Refinement Module (ASTRM) on the RegNet-Y backbone. In addition, the Soft Instance Contrastive (SoftIC) Loss function is specifically applied to address the class imbalance problem. The research methodology designs five test scenarios to evaluate the effectiveness of the integration of ASTRM and SoftIC Loss compared to the baseline model. Model performance is measured using the mean Average Precision (mAP) metric with a temporal tolerance of 1 second according to the SoccerNet evaluation protocol. Experiment result in the best scenario successfully improved the mAP performance to 57,75% from the T-DEED baseline performance of 53% mAP. The results of this study are expected to contribute to improving the accuracy of soccer action spotting, especially in class imbalance conditions, and to enrich the study of sports video-based action spotting, especially in football video.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Action Spotting, ASTRM, Deep Learning, SoccerNet, SoftIC Loss, T-DEED
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Ivan Fairuz Adinata
Date Deposited:	23 Jul 2026 03:13
Last Modified:	23 Jul 2026 03:13
URI:	http://repository.its.ac.id/id/eprint/137099

Actions (login required)

View Item