Action Detection in Badminton Courts Using AVA Dataset and MMAction2 Architecture with Slow Fast Model

Abstract

As a matter of fact, action detection technologies have emerged as powerful tools for a multitude of applications, from surveillance and security to healthcare and sports analytics. Badminton courts, especially those that are highly frequented, present unique challenges owing to the concentration of people engaged in diverse activities. Incidents such as unintended collisions between players, or unauthorized walking across the courts, are not uncommon and necessitate efficient monitoring for risk mitigation. This study addresses these issues by employing the MMAction2 architecture and the Slow Fast model for action detection in badminton courts. The author uses a dataset collected from multiple badminton facilities and leverages the AVA dataset for training and validation. The results are promising, with the model showing high levels of accuracy in identifying various types of actions: playing badminton, sitting, walking across the court, falling, and watching the game. The implications of this research are significant for badminton court management and safety.