Unified adaptive open-set incremental learning for industrial human action recognition

The growing demands of intelligent manufacturing call for robust and efficient systems for industrial human action recognition (HAR). However, existing closed-set HAR methods struggle with novel actions and lack adaptability in dynamic industrial scenarios. Aiming at overcoming these limitations, a unified adaptive approach for open-set incremental learning was proposed, extending conventional action recognition to handle novel categories and evolving tasks. Specifically, angular prototype learning (APL) constructs a compact and well-separated feature space, within which deep deterministic uncertainty (DDU) models class-conditional feature distributions to reliably detect previously unseen actions. Detected unknowns are organized through unsupervised hierarchical clustering with fused deep semantic and edge features, reducing annotation effort while an APL-consistent replay-based incremental learning (IL) facilitates sequential incorporation of novel actions while mitigating forgetting. Experiments on datasets collected from real-world manufacturing workshops validated the approach's effectiveness and adaptability across different HAR architectures, demonstrating reduced annotation effort and improved long-term stability in dynamic industrial scenarios.