Human Activity Recognition Using Self-Powered Sensors Based on Multilayer Bidirectional Long Short-Term Memory Networks

Sensor-based human activity recognition (HAR) requires the acquisition of channel state information (CSI) data with time series based on sensors to predict human behavior. Many existing approaches are based on wearable sensors and cameras, which increases the burden and privacy issues for patients. Self-powered sensors are capable of noncontact collection of time series data generated by human activity while ensuring their own stable operation. In this article, we propose a deep-learning-based framework for contactless real-time activity detection of humans using self-powered sensors, which is called multilayer bidirectional long short-term memory (MBLSTM). The collected Wi-Fi CSI data are fed into our proposed network model, which is then used to learn representative features of both sides from the original continuous CSI measurements. The attention model is used to assign different weights to the learned features, and finally, activity recognition is performed. Experimental results show that our proposed method achieves an accuracy of more than 96% for the recognition of six activities in multiple rounds of testing, outperforming other benchmark methods used for comparison.