Lightweight Action Recognition Based on Selective Self Training and Hot Swapping Mechanism

The Naive Bayes algorithm, while widely used in action recognition tasks, faces limitations in its generalization ability due to the assumption of feature independence. Despite these challenges, it remains a valuable tool in scenarios with constrained computational resources, owing to its low complexity and robustness against noise. To address these contradictions, this paper proposes a lightweight action recognition framework that integrates intelligent sample filtering, self-training methods, and a hot swapping mechanism. By employing a sample filtering mechanism, the quality of the training data is significantly improved, ensuring more accurate model predictions. Additionally, the self-training approach facilitates application-side training, thereby enhancing the model's personalized adaptability to different environments and users. The hot swapping mechanism enables seamless replacement of the model in real-time, ensuring that the system can dynamically adjust to new data without disrupting the user experience. The experimental results demonstrate the effectiveness of the proposed framework, with the project team successfully completing the entire process from training to recognition. A test set, representing 20% of the total data, was separated from user-recorded training samples. The framework achieved an accuracy rate exceeding 90%, proving its efficiency and reliability in real-world applications.