Deep Learning for Device-free Human Activity Recognition Using WiFi Signals

WiFi-based device-free human activity recognition applications have been popular for a decade in smart-environment sensing domains. Existing researches extract features of human activity as the input of classic classification methods to train the learning model and then leverage the trained model to recognize activities. The propagation characteristics of WiFi signals are variable for individuals under different place conditions even in the same environment with time. Depending on fixed features, data cannot effectively represent human activity and adapt to a dynamic indoor environment. As deep learning has demonstrated its effectiveness in device-free sensing domains in recent years, deep learning methods have been explored to address difficulties faced by WiFi signals. In this chapter, we focus on how to weaken the accuracy differences among individuals on human activity recognition and improve the robustness in a single indoor environment. Based on this, we design a novel deep learning model called LCED which consists of one LSTM-based Encoder, features image representation and one CNN-based Decoder. LSTM-based Encoder can learn time-sequence representation and encode it to an equal-length vector. Each equal-length vector is represented using features of image representation to compress and keep key details. CNN-based Decoder provides better recognition accuracy by capturing the local effective information of the signal image based on the spatial distribution. Experimental results show that the average accuracy of sixteen activities is high, 95%. The accuracy difference among individuals on activity recognition averagely decreases by 3%.