ExtRe: Extended Temporal-Spatial Network for Consumer-Electronic WiFi-Based Human Activity Recognition

Abstract

With consumer electronics (CE) development, consumer-electronic WiFi-based human activity recognition (HAR) has been acknowledged as an essential non-intrusive technology in several crucial human-oriented fields, such as metaverse accessing, sign language communication, and healthcare monitoring. By learning from the experience from the CV field, i.e., treating wireless signals as images and extracting the feature based on a 2-dimensional convolutional neural network (2D CNN), existing studies have achieved considerable progress. However, there is a critical difference in feature compositions between WiFi signals and images. Whether existing methods apply to coarse-grained or fine-grained activity recognition, their performance is limited due to the loss of important dynamic temporal features hidden among multiple channels. To overcome these problems, we propose an extended temporal-spatial approach for WiFi-based HAR, named ExtRe, in which sufficient attention is paid to both temporal and spatial-channel characteristics. We also considered a challenging case, i.e., coarse-grained or fine-grained activities. Experiment results based on six datasets show that, compared with the state-of-the-art methods, our ExtRe achieves superior performance. The proposed ExtRe achieves 100% accuracy on a fine-grained dataset. On the coarse-grained dataset, ExtRe improves the accuracy by 0.6% with high stability. In addition, ExtRe has 20.67% less floating point operations (FLOPs).