Cross-subject human activity recognition based on self-training and self-supervised learning Cross-Subject Human Activity Recognition based on Self-training

• Published in: Computing Vol. 107; no. 6
• Main Authors: Zhang, Qi, Wei, Baichun, Yi, Chunzhi, Zhu, Haiqi, Ding, Qiang, Jiang, Feng
• Format: Journal Article
• Language: English
• Published: Vienna Springer Vienna 01.06.2025
• Subjects: Artificial Intelligence; Computer Appl. in Administrative Data Processing; Computer Communication Networks; Computer Science; Information Systems Applications (incl.Internet); Regular Paper; Software Engineering; 68T01; Cross-subject adaptation; Human activity recognition; Unsupervised learning; Wearable sensors
• ISSN: 0010-485X; 1436-5057
• DOI: 10.1007/s00607-025-01503-1

Human activity recognition (HAR) holds promise for applications in healthcare and smart home systems. However, a notable reduction in recognition accuracy stemming from the cross-subject issue significantly impedes the widespread implementation of HAR. To address this issue, we introduce an unsupervised method that utilizes domain-level distribution commonality and sample-level data similarity to adapt a HAR model to the target domain (i.e., new subjects). Our method comprises three steps. Firstly, we pre-train a HAR model in the source domain and employ it to select samples in the target domain by a predefined confidence threshold. These samples are then merged with the source domain, constructing a cross-domain labeled dataset. Secondly, we design a neighborhood clustering loss considering spatiotemporal correlations among samples within the target domain. The loss is employed to cluster each sample’s neighbors in a self-supervised manner. Thirdly, we update the pre-trained model by multi-mask learning. The cross-entropy loss and the neighborhood clustering loss are applied on the cross-domain dataset and the training set of the target domain, respectively. We evaluate the proposed method on three public datasets by Leave-One-Subject-Out Cross-Validation (LOSO-CV). The method achieves state-of-the-art performance with average accuracies of 94.14%, 88.99% and 87.95% on these datasets, respectively. Our method is characterized by its user-friendliness and holds promising prospects for applications in health services and in-home monitoring systems.