Detection of tennis ball landing points in videos using enhanced ppTSM v2 recognition model

The conventional method of manually marking the landing points of tennis balls is inefficient, labor-intensive and prone to subjective errors, which limits accuracy and efficiency. This study aims to develop a video-based tennis landing point detection system that allows players to analyze and refine their stroke preferences, while providing coaches and referees with more insightful data support. We have introduced a new tennis landing point dataset, T-Point, which includes different tennis usage scenarios. The PP-TSMv2 model, which uses a 2D network, was chosen as the recognition framework. To improve the performance of the model in recognizing tennis landing points, we replaced the SE module in its PP-LCNetV backbone network with the convolutional attention module CBAM. The original model achieved a top1_avg of 0.6889 on the T-Point dataset, while the refined model achieved a top1_avg of 0.7333. The experimental results confirm that the integration of spatial information significantly improves the performance of the model.