Federated learning based personalized cross domain martial arts action recognition model

Martial arts action recognition plays a pivotal role in cultural heritage preservation, intelligent sports training, and human behavior analysis. However, traditional centralized learning paradigms are challenged by data privacy leakage and cross-domain distribution shifts. To address these issues, this paper proposes a cross-domain framework named Federated Personalized Domain Adaptation for Martial Arts Action Recognition. Without sharing raw skeleton sequence data, the framework utilizes a dynamic domain alignment module to extract global shared features, while introducing client-side personalized adaptation layers to mitigate domain discrepancies arising from diverse martial arts styles and acquisition devices. Extensive experiments conducted on the self-constructed CMA-2025 dataset and the public NTU RGB + D120 dataset demonstrate that the proposed method achieves an average recognition accuracy of 94.8% in cross-domain scenarios, outperforming baseline methods by 7.3%, while exhibiting significant advantages in privacy protection metrics. This study provides an efficient and trustworthy technical paradigm for the digital preservation of martial arts.