A Real-time Hybrid Method for Basketball Key Actor Analysis and Event Detection

In the present era, teams leverage technology to analyze events and key players in a basketball game, often beneficial for information gathering and tactical analysis. However, existing methods for event detection and key player identification often struggle with real-time execution and require a huge amount of training data. This paper introduces a hybrid system that combines rule-based methods with deep learning models to address this issue. Our system demonstrates strong generalization ability across different basketball courts on three tasks: temporal action localization, spatio-temporal action detection, and key actor detection. We evaluate the proposed system on the MultiSports dataset and two self-collected datasets: the P.League+ Basketball Event (P-Event) dataset for validating action localization and the P.League+ Key Actor (P-KeyActor) dataset for validating key actor detection. Our method achieves a 55.5% F1 score on the event localization task (on P-Event) and 81.5% on the key actor detection task (P-KeyActor) with real-time (30 FPS) performance. Our system is modular; improvements to any single module can significantly enhance overall system performance.