Accurate Calibration of Physical Education Teaching Action Based on Artificial Intelligence Deep Learning Technology

Traditional manual scoring of physical education is limited by subjectivity and its inability to capture high-frequency, subtle differences in movement. This article constructs a technical route of multi-source perception, feature fusion, accurate calibration, and closed-loop feedback. To guarantee the space of each frame is consistent, a timestamp alignment method and skeleton relocation strategy are designed. Subsequently, the study proposes a multi-scale skeleton-feature fusion network. The local topology is extracted by gated convolution, while global dynamics are integrated through a cross-layer attention mechanism. The model is jointly optimized using joint position loss and motion trajectory loss. This study verifies the feasibility of integrating deep learning with multi-sensor data for the accurate calibration of physical education actions. The study provides a data-driven basis for classroom hierarchical error correction, action prescription, and early warning of sports injuries. It also has direct guiding significance for building real-time intelligent teaching systems.