Domain generalization with semi-supervised learning for people-centric activity recognition  

作　　者：Jing LIU Wei ZHU Di LI Xing HU Liang SONG 

机构地区：[1]Academy for Engineering&Technology,Fudan University,Shanghai 200433,China [2]Shanghai East-bund Research Institute on Networking Systems of AI,Shanghai 202162,China [3]School of Optoelectronic Information and Computer Engineering,University of Shanghai for Science&Technology,Shanghai 200093,China

出　　处：《Science China(Information Sciences)》2025年第1期168-185,共18页中国科学(信息科学)(英文版)

基　　金：supported in part by China Mobile Research Fund of the Chinese Ministry of Education(Grant No.KEH2310029);Specific Research Fund of the Innovation Platform for Academicians of Hainan Province(Grant No.YSPTZX202314);supported by the Shanghai Key Research Laboratory of NSAI and the Joint Laboratory on Networked AI Edge Computing Fudan University-Changan.

摘　　要：People-centric activity recognition is one of the most critical technologies in a wide range of real-world applications,including intelligent transportation systems,healthcare services,and brain-computer interfaces.Large-scale data collection and annotation make the application of machine learning algorithms prohibitively expensive when adapting to new tasks.One way of circumventing this limitation is to train the model in a semi-supervised learning manner that utilizes a percentage of unlabeled data to reduce the labeling burden in prediction tasks.Despite their appeal,these models often assume that labeled and unlabeled data come from similar distributions,which leads to the domain shift problem caused by the presence of distribution gaps.To address these limitations,we propose herein a novel method for people-centric activity recognition,called domain generalization with semi-supervised learning(DGSSL),that effectively enhances the representation learning and domain alignment capabilities of a model.We first design a new autoregressive discriminator for adversarial training between unlabeled and labeled source domains,extracting domain-specific features to reduce the distribution gaps.Second,we introduce two reconstruction tasks to capture the task-specific features to avoid losing information related to representation learning while maintaining task-specific consistency.Finally,benefiting from the collaborative optimization of these two tasks,the model can accurately predict both the domain and category labels of the source domains for the classification task.We conduct extensive experiments on three real-world sensing datasets.The experimental results show that DGSSL surpasses the three state-of-the-art methods with better performance and generalization.

关 键 词：activity recognition deep learning domain generalization semi-supervised learning adversarial training 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]