Measuring intrinsic human activity information using WiFi-based attention model.

[Display omitted] • An attentional deep learning model, named DeepFocus, is proposed to allow selective retrieval of informative signal variations and crucial feature regions. • An innovative temporal attention module based on the structural similarity values of WiFi snippets is designed. • A channel-spatial attention module is presented to strengthen the useful signal features along both the channel and spatial axes. • Extensive experiments are carried out on various real activity datasets. In this work, we have proposed the first WiFi-based attention model for intrinsic human activity information measuring. First, both convolutional layers and time-recurrent layers are integrated in a one-off manner for joint feature learning. Based on this, a temporal attention module is introduced to capture activity moments and reduce lengthy data collections. In addition, a channel-spatial attention module is further designed to retrieve the activity feature of interest, without requiring any additional supervision. Extensive experiments were conducted for performance comparison under diverse environmental settings. The results showed that our model can achieve the best overall accuracy of 94.62%, 94.36% and 91.04% in three datasets with minimal manual efforts involved, and consistently outperforms other baselines under diverse parameter settings. [ABSTRACT FROM AUTHOR]