Real-time table tennis detection and motion trajectory forecasting using improved YOLOv8 and GRU

To enhance the perception and prediction performance of table tennis perception and future movement, this study presents a framework for detecting and predicting the future movement of the table tennis game based on an optimized version of YOLOv8 and the Gated Recurrent Unit. On the detection side, this study designs a detail-preserving and motion-consistency-enhanced YOLOv8. The network introduces a P2 high-resolution branch, spatial rearrangement downsampling, lightweight cross-scale attention, and motion-consistency confidence recalibration to strengthen feature representation and interference suppression for high-speed small targets. On the forecasting side, this study develops a Physics-constrained Transformer-Gated Recurrent Unit network. This network uses constant acceleration Kalman filter algorithm, time-based feature learning, residual correction mechanism, and bounce-on-table constraint for predicting future trajectories. The experiment result shows that using an improved detection network yields an F1 score of 95.36% under blurred and occluded scenarios. The normalized ball-center error decreases to 0.52%, and the inference speed reaches 151.98 FPS. For motion trajectory forecasting, the average displacement error within a 300 ms interval is 44.91 mm, while the landing-point error is 50.68 mm. The physical violation rate is only 3.12%, and the recovery rate under controlled missing observations reaches 90.61 ± 1.05%. These findings prove that the introduced model is able to provide significantly more robust high-speed detection of table tennis as well as more physically realistic long-term trajectory forecasting. The model is thus able to offer reliable technical assistance in intelligent training and match analysis. Not applicable.