Computational Analysis of Table Tennis Games from Real-Time Videos Using Deep Learning
Utilizing a multiscale training dataset, YOLOv8 leverages deep learning to deliver rapid inference capabilities and exceptional accuracy in detecting visual objects, particularly smaller ones. This performance surpasses that of transformer-based deep learning models, positioning YOLOv8 as a leading algorithm in its field. While the effectiveness of visual object detection is generally assessed using pre-trained models on enhanced datasets, fine-tuning becomes crucial for specific situations like table tennis matches and coaching sessions. The unique challenges in these contexts include rapid ball movement, uniform color, fluctuating lighting conditions, and bright reflections due to intense illumination. In this thesis, we introduce a motion-centric algorithm to augment the YOLOv8 model, aiming to improve the accuracy of predicting ball trajectories, impact points, and velocity in the realm of table tennis. This adaptive model not only elevates its utility in real-time sports coaching but also demonstrates its potential in other fast-paced settings. Experimental results indicate a significant improvement in detection rates and a reduction in false positives.