CILOSR: A unified framework for enhanced class incremental learning based open-set human activity recognition using wearable sensors

Abstract

The field of Human Activity Recognition (HAR) has seen widespread adoption of wearable sensors for the collection of time-series signals. However, as new activities emerge, HAR systems struggle to differentiate novel categories from existing ones, as they are trained on a fixed set of known classes. To overcome this limitation, an innovative framework called CILOSR is designed for the continuous integration of novel, previously unseen activity classes into HAR models. The proposed CILOSR framework combines two pivotal processes, Class Incremental Learning (CIL) to enhance model knowledge with newly acquired data, while Open-Set Recognition (OSR) to detect and characterize new activity classes. The CIL phase employs extreme point updating based Extreme Value Machine algorithm, which preserves and updates the reference boundary points and extreme value vectors for established classes alongside new data integration. For the OSR phase, Principal Component Analysis (PCA) is incorporate to reduce feature redundancy within the time–frequency domain, thereby refining the feature space. Subsequently, Particle Swarm Optimization (PSO) is utilized for precise calibration of Extreme Value Machine (EVM) parameters to optimize the recognition process. Several experiments on the UCI, PAMAP2, and USC-HAD datasets confirm the effectiveness of the CILOSR framework. Specifically, OSR-LPC (Leave-Partial-Class) experiments on the UCI dataset demonstrate that CILOSR with PSO-EVM (Cosine) + PCA significantly outperforms the standard EVM (Cosine). The model achieves F1-macro score of 0.88 and accuracy of 0.89, compared to the baseline’s 0.59 and 0.66. These results highlight CILOSR’s enhanced accuracy in recognizing both known and unknown activities, demonstrating its potential for dynamic and scalable HAR applications.