A Deep Learning Framework for Joint Activity Recognition and Indoor Localization Using Wi-Fi CSI

In recent years, With the use of deep learning in Wi-Fi smart sensing, the efficiency of non-contact sensing technology has been greatly improved, such as heartbeat monitoring, indoor positioning, activity recognition and so on. This paper introduces a deep learning framework that combines the SE1D attention mechanism module, the Xception Time model, and the Automatic Weighted Loss to implement the Joint Activity Recognition and Indoor Localization (JARIL) task using Wi-Fi CSI data. The framework explicitly models the dependencies between channels through the SE1D attention mechanism module, which effectively improves the adaptability and sensitivity of the network to the importance of different channel features. The lightweight Xception Time model is used to improve the generalization ability. Automatic Weighted Loss is used to dynamically adjust the weight of the loss function during the training process to reduce the training cost. The experimental results show that the highest accuracy of activity recognition of the proposed method is 91.96%, and the positioning accuracy is 100%. Compared with other advanced deep learning models, it has better performance.