Automatic Detection of Soccer Events using Game Audio and Large Language Models
Master thesis
Permanent lenke
https://hdl.handle.net/11250/3147993Utgivelsesdato
2024Metadata
Vis full innførselSamlinger
- Master's theses (RealTek) [1826]
Sammendrag
This thesis tackles the inefficiencies associated with manual annotation in soccer event detection, a process that is time-consuming, expensive, and difficult to scale during major tournaments. By developing an automated audio-based event detection system, this research aims to bypass the extensive resource requirements of traditional video action detection, offering a more efficient and balanced alternative. This research uses the Automatic Speech Recognition (ASR) transcriptions from SoccerNet-Echoes to contribute with two supervised datasets: the 15-Second Standard Deviated Dataset (15-SSDD) and 30-Second Standard Deviated Dataset (30-SSDD). These datasets incorporate a 15-second and 30-second standard deviated window of soccer event context to train models for recognising key events like Goals, Fouls, and Corners. They were evaluated on several Large Language Models (LLMs, including DistilBERT, BERT BASE, BERT LARGE, and all-MiniLM-L6-v2. The findings show that longer contextual samples significantly enhance the model's classification accuracy, underscoring the importance of context within events in soccer. The all-MiniLM-L6-v2 model is noted for its high accuracy and computational efficiency, making it ideal for real-world applications that demand rapid and precise event detection. It performs robustly across various metrics such as F1-score, precision, and recall. It operates efficiently on both datasets with fewer computational resources, underscoring its suitability for efficient and accurate applications. Challenges such as class imbalance impact the overall effectiveness of the detection system. The thesis proposes future enhancements, including audio implementation and exploring class balancing strategies like word embedding oversampling and cost-sensitive learning to refine the system's robustness and effectiveness. This research advances the field of sports analytics by proposing an efficient audio-based event detection system. It also sets the stage for future innovations that could transform the monitoring and analysis of sports events, enhancing viewer experiences and providing sports professionals with critical insights in real-time.