Résumé: Human Activity Recognition (HAR) based on Channel State Information (CSI) has important application prospects in various fields such as human–computer interactivity, medical health et al. Although the current CSI-based HAR researches have made great progress in the number of activity categories and recognition accuracy, they encounter two challenges. When recognizing new activities, a large number of new activity samples are required to retrain the original model or adjust the model parameters. Besides, the recognition accuracy of the model retrained in the new scene for the already recognized activities drops obviously. To address the two challenges, we propose a human activity recognition system CSI-ARIL based on incremental learning in this paper. CSI-ARIL saves partial representative samples for each learned activity, and adds distillation loss to the loss function as the regularization term to further strengthen the model’s memory for the learned activities. In order to enhance the feature difference among different activities, CSI-ARIL explicitly adds time information for CSI activity samples. In addition, pseudo-samples are generated through data enhancement to expand the new training set, which reduces the cost of obtaining labeled real samples. CSI-ARIL also adds the Convolutional Block Attention Module (CBAM) to the network, which helps the network extract the effective features of CSI activity samples. The experimental results show that, after continuously learning eight categories of activities in four scenes, the recognition accuracy of CSI-ARIL for new and already recognized activities reaches 92.8% and 89.4%, respectively.

Langue: Anglais