Computational Analysis of Table Tennis Matches from Real-Time Videos Using Deep Learning

Abstract

In this paper, utilizing a multiscale training dataset, YOLOv8 demonstrates rapid inference capabilities and exceptional accuracy in detecting visual objects, particularly smaller ones. This outperforms transformer-based deep learning models, makes it a leading algorithm in its domain. Typically, the efficacy of visual object detection is gauged by using pre-trained models based on augmented datasets. Yet, for specific situations like table tennis matches and coaching sessions, fine-tuning is essential. Challenges in these scenarios include the rapid ball movement, color, light conditions, and bright reflections caused by intense illumination. In this paper, we introduce a motion-centric algorithm to the YOLOv8 model, aiming to boost the accuracy in predicting ball trajectories, landing spots, and ball velocity within the context of table tennis. Our adapted model not only enhances the real-time applications in sports coaching but also showcases potential for applications in other fast-paced environments. The experimental results indicate an improvement in detection rates and reduced false positives.