Fast detection, localization, and trajectory prediction of table tennis based on shuffle YOLOv5s model

With the rapid development of deep learning and computer vision technologies, target detection and trajectory analysis in ball sports have become an important research direction in intelligent sports. However, existing table tennis detection methods generally suffer from issues such as insufficient detection accuracy and poor robustness in complex game environments. To address this, this study proposes an integrated model for fast detection, localization, and trajectory prediction tailored to table tennis game scenarios. First, a lightweight and high-precision object detection network is constructed based on the You Only Look Once version 5 (YOLOv5) model, combined with the ShuffleNetv2 backbone structure and the attention mechanism module. Subsequently, a time-series fusion model is designed for accurate prediction of table tennis ball trajectories by combining gated recursive units with long and short-term memory networks. Experimental results show that the proposed detection model achieves accuracy, recall, and F1 scores of 0.982, 0.964, and 0.971, respectively, while the trajectory prediction module achieves prediction accuracy rates of 98.2% and 98.6% on the training and testing datasets, respectively. The research results validate the effectiveness and practical value of this method in intelligent detection and trajectory tracking tasks for table tennis, providing technical support for future intelligent sports analysis systems.