Abstract
BACKGROUND: With the deepening implementation of the “Physical Education and Teaching Integration” strategy, the role of physical education in the modern educational system has been increasingly emphasized. In the context of the “Double Reduction” policy, the transformation toward personalized, scientific, and information-driven teaching has become a key challenge. METHODS: This paper proposes a personalized instructional support system for physical education based on multimodal artificial intelligence technologies. The system is designed to deliver precise instructional interventions through action recognition, error analysis, and stratified feedback mechanisms.The system integrates video imagery and skeletal pose data, adopting Video MAE V2 as the core recognition framework. To enhance inference efficiency and real-time responsiveness, progressive knowledge distillation (PKD) and weighted inference with scaled ensemble (WISE) are introduced. Furthermore, a personalized feedback mechanism is developed that analyzes error vectors and generates stratified training recommendations tailored to individual students, thereby promoting differentiated instruction. RESULTS: The proposed method achieves excellent performance on the UCF101 and HMDB51 datasets. Additionally, the model exhibits strong engineering applicability, maintaining efficient operation on edge computing platforms such as Raspberry Pi and Jetson Nano. CONCLUSION: The system holds promise for scalable applications in intelligent physical education. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s13102-026-01574-3.