A Lightweight Ensemble Framework for Online Skeleton-Based Human Action Recognition in Industrial Environments

Online human action recognition in industrial environments poses unique challenges, as it requires accurate, real-time identification of ongoing actions under diverse and often unpredictable operating conditions. In this work, we propose a simple yet effective framework designed to address these challenges through a combination of lightweight, adaptable components. The framework integrates (i) an ensemble of binary classifiers, enabling flexible and application-specific action detection with minimal training data, and (ii) a sliding window strategy including per-frame preprocessing to support real-time inference on continuous data streams. To improve generalization across different environments, our approach relies on skeleton data, which provide a compact yet informative representation of human movements and are inherently more robust to variations in viewpoint and background than video-based methods. The proposed framework was validated in both a laboratory and a real industrial setting, considering a collaborative carbon fiber draping task. Experimental results showed that the proposed framework achieves accurate and timely recognition of relevant human actions, while demonstrating strong generalisability to different scenarios with minimal training data, making it a practical and efficient solution for real industrial processes.