Optimization method for catching deviation of table tennis training robot based on physical motion model and YOLOv3

To solve the receiving deviation caused by inaccurate visual detection, large trajectory prediction errors, and system delay in table tennis training robots, the core idea of "perception-prediction-control" full chain collaborative optimization is taken to improve system performance. Firstly, the detection network You Only Look Once version 3 (YOLOv3) is improved by embedding convolutional attention modules, adaptively spatial feature fusion techniques, and CIoU loss to optimize the accuracy of small object detection. Then, a high-order physical motion model integrating spin and aerodynamics is constructed, combined with extended Kalman filter to achieve optimal trajectory estimation. Finally, a composite control strategy combining feedforward trajectory planning and feedback compensation is designed to counteract system delay. The improved detection model had an accuracy of 98.8% ± 0.5%, a recall of 97.5% ± 0.6%, and an average accuracy of 98.2% ± 0.4%. The average trajectory prediction error of the extended Kalman filter fused with high-order models was as low as 8.7 mm ± 1.2 mm, which was 75.6% lower than that of the ideal projectile model. The feedforward + feedback control strategy reduced the median deviation of hitting to 12.4 mm, the catching success rate reached 95.2%, and reduced the total system delay from 42ms to 32ms. The conclusion indicates that multi-module collaborative optimization can significantly reduce the catching deviation of table tennis training robots. The research provides technical support for the intelligent upgrade of table tennis training robots, and also provides reference for motion target tracking and interception devices.