A Study on Hit Detection in Tennis Based on Data from Diverse Environments Using Transformer

Abstract

Currently, data for tactical analysis in tennis is manually annotated, which is a heavy burden. Therefore, this study aims to perform hit detection that accurately predicts the hit frame from tennis videos. Previous studies have used the same environment for training and test data, and have not adapted to videos in unknown environments. For practicality, this study considers improving the performance of hit detection in unknown environments by training with data from diverse environments. Since the Transformer has the characteristic of improving accuracy when using large-scale data, we hypothesized that the Transformer could be used to solve the above problem. We propose using a Transformer-based ActionFormer as a model for this study. In our experiments, we created models for the players at the front and back of the tennis court, and evaluated the performance of hit detection in the video using F1 values. The F1 values of the proposed method were 0.762 and 0.738, which were more than 10% more accurate than 0.636 and 0.538 for the near and far players, respectively, in the previous study.