In competitive table tennis training, accurate identification and analysis of ball landing points are essential for tactical and technical improvement.This study presents a deep learning-based system integrating high-speed visual perception, three-dimensional trajectory reconstruction, and landing area classification to provide data-driven training optimisation.The core novelty lies in fusing multi-frame time-series modelling with physical rebound constraints and personalised training prescriptions driven by weighted KL divergence.Major technical contributions include small-target enhancement, motion blur compensation, and regional classification error correction.The evaluation protocol involved 418 recorded rounds with high-speed multi-camera setups, achieving a touchdown recognition accuracy of 0.931, regional classification consistency of 0.927, and boundary misjudgement reduction from 0.083 to 0.039.Compared to baseline methods, coverage of deep and corner areas increased from 0.287 to 0.361 and 0.174 to 0.248 respectively, and distribution entropy improved from 2.91 to 3.38, demonstrating substantial quantitative improvements in tactical diversity and training effectiveness.
Zheming Liu (Thu,) studied this question.