Multi-Object Tracking and Temporal Action Localization for Basketball Video Understanding
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Significant advances in deep learning-based object detection have led to substantial progress, such as in Multi-Object Tracking (MOT) and Temporal Action Localization (TAL) algorithms.However, current methods mainly foc us on pedestrian or vehicle scenes, while team-sports scenarios, such as i n basketball, present unique challenges, including three or more object oc clusions with similar appearances and high-intensity complex motions, refe rred to as Complex Multi-Object Occlusion (CMOO) problems.This thesis add resses fundamental challenges in computer vision for basketball video anal ysis through two complementary contributions; a specialized multi-object t racking algorithm and a comprehensive temporal action localization framewo rk.As the first study, Basketball Simple Online and Realtime Tracking (Bask etball-SORT), an online and robust MOT method specifically designed to sol ve CMOO problems in basketball videos, is proposed.Instead of the traditi onal Intersection-over-Union (IoU) based approach, trajectories of neighbo ring frames based on projected positions of players are utilized.Two doma in-specific mechanisms are designed: a constraint mechanism that leverages the spatial structure of basketball courts and Basketball Game Restrictio n (BGR) to reduce tracking ambiguity, and a re-identification mechanism ca lled Reacquiring Long-Lost ID (RLLI) for handling occlusions.These mechan isms, combined with player trajectory analysis and appearance features, ef fectively solve occlusion problems.Experimental validation on fixed-camer a basketball datasets demonstrates that 63.48\% Higher Order Tracking Accu racy (HOTA) on 5x5 basketball and 66.45\% HOTA on 3x3 basketball are achie ved, substantially outperforming state-of-the-art methods and demonstratin g robustness across different basketball game formats.
Fields of Study & Subjects
Video Surveillance and Tracking Methods
Human Pose and Action Recognition
Advanced Neural Network Applications
Publication details
- Journal / source
- Nagoya University Repository (NAGOYA Repository)
- Publication date
- 30 January 2026
- Language
- EN
- Document type
- Thesis