Implementasi Tracking Berbasis Deep Learning untuk Multi-Objek Pemain Sepak Bola pada Dataset Soccernet

Maheswara, Stevanza Gian (2026) Implementasi Tracking Berbasis Deep Learning untuk Multi-Objek Pemain Sepak Bola pada Dataset Soccernet. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221248-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (11MB) | Request a copy

Abstract

Analisis performa sepak bola modern membutuhkan data pelacakan pemain yang akurat, namun tracking multi-objek pada video pertandingan masih menghadapi tantangan yang berat. Pergerakan pemain yang cepat dan non-linear, oklusi antar-pemain, gerak serta zoom kamera siaran, dan terutama kemiripan visual akibat seragam satu tim yang identik menyebabkan sistem rentan kehilangan jejak maupun mengalami pertukaran identitas (ID switching). Tantangan inilah yang membuat konsistensi identitas pemain sulit dipertahankan sepanjang sekuens video sepak bola.
Penelitian ini mengimplementasikan tracking berbasis deep learning untuk multi-objek pemain sepak bola pada dataset SoccerNet, dengan Deep OC-SORT sebagai metode yang diusulkan dan lima metode pembanding, ByteTrack, OC-SORT, StrongSORT, DeepSORT, dan FairMOT. Deep OC-SORT menyempurnakan OC-SORT dengan mengintegrasikan fitur visual (Deep Re-Identification), Kompensasi Gerak Kamera (CMC), dan Adaptive Weighting. Untuk menjamin perbandingan yang adil, seluruh metode tracking-by-detection menggunakan detektor yang sama, yaitu YOLOX-M yang di-fine-tune pada dataset SoccerNet, sehingga perbedaan performa dapat difokuskan terutama pada kualitas asosiasi. Seluruh metode dievaluasi pada data uji SoccerNet Tracking (49 sekuens, SNMOT-116 hingga SNMOT-200) menggunakan library TrackEval dengan metrik HOTA, DetA, AssA, MOTA, IDF1, dan IDSW, serta dilengkapi studi ablasi untuk mengukur kontribusi tiap komponen.
Hasil eksperimen menunjukkan Deep OC-SORT memperoleh performa terbaik dengan HOTA 62,84%, AssA 54,97%, IDF1 72,07%, dan jumlah pertukaran identitas terendah (IDSW 1.899), didukung detektor YOLOX-M yang kuat (AP@0,50 sebesar 94,6%). Dibandingkan basisnya, OC-SORT, Deep OC-SORT menaikkan AssA 2,01 poin, IDF1 2,10 poin, serta menurunkan IDSW sekitar 38%. Studi ablasi mengungkap bahwa peningkatan ini didapatkan dari modul CMC, sedangkan kontribusi (Deep ReID) dan Adaptive Weighting tidak begitu signifikan. Temuan ini menjelaskan bahwa pada domain sepak bola dengan visual yang identik, estimasi pergerakan kamera berperan jauh lebih besar daripada fitur visual dalam menjaga konsistensi identitas pemain.
================================================================================================================================
Modern soccer performance analysis relies on accurate player tracking data, yet multi-object tracking in match footage still faces considerable challenges. Fast, non-linear player movement, frequent occlusion between players, broadcast camera panning and zoom, and most critically high visual similarity caused by identical team jerseys make the system prone to losing tracks and to identity switches (ID switching). These challenges make it difficult to maintain consistent player identity throughout a soccer video sequence. This research implements deep learning-based tracking for multi-object soccer players on the SoccerNet dataset, with Deep OC-SORT as the proposed method and five baselines, ByteTrack, OC-SORT, StrongSORT, DeepSORT, and FairMOT. Deep OC-SORT enhances OC-SORT by integrating appearance features (Deep Re-Identification), Camera Motion Compensation (CMC), and Adaptive Weighting. To ensure a fair comparison, all tracking-by-detection methods share the same detector YOLOX-M fine-tuned on SoccerNet so that performance differences can be focused primarily to association quality. All methods are evaluated on the SoccerNet Tracking test set (49 sequences, SNMOT-116 to SNMOT-200) using the TrackEval framework with HOTA, DetA, AssA, MOTA, IDF1, and IDSW metrics, complemented by an ablation study to measure the contribution of each component. The experimental results show that Deep OC-SORT achieves the best performance, with 62.84% HOTA, 54.97% AssA, 72.07% IDF1, and the lowest number of identity switches (IDSW 1,899), supported by a strong YOLOX-M detector (94.6% AP@0.50). Compared to its baseline OC-SORT, Deep OC-SORT improves AssA by 2.01 points and IDF1 by 2.10 points, and reduces IDSW by approximately 38%. The ablation study reveals that this improvement is driven primarily by the CMC module, while the contributions of appearance features (Deep ReID) and Adaptive Weighting are nearly negligible. These findings confirm that in the soccer domain with identical jerseys, camera motion estimation plays a far greater role than appearance features in maintaining player identity consistency.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Kompensasi Gerak Kamera, Deep OC-SORT, Pelacakan, Multi-Objek, SoccerNet, Sepak Bola. Camera Motion Compensation, Deep OC-SORT, Multi-Object Tracking, SoccerNet, Soccer.
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing. T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK6592.A9 Automatic tracking.
Divisions:	Faculty of Information Technology > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Stevanza Gian Maheswara
Date Deposited:	27 Jul 2026 06:27
Last Modified:	27 Jul 2026 06:27
URI:	http://repository.its.ac.id/id/eprint/137842

Actions (login required)

View Item