Device-free Location-independent Human Activity Recognition using Transfer Learning based on CNN

Device-free human activity recognition based on wireless signal is becoming a vital underpinning for various emerging applications in human-computer interaction (HCI). Ubiquitous wireless communication network, especially WiFi promotes the development of relevant industrial applications as well as the academic researches. Without dedicated equipment and specific constraints, device-free human activity sensing based on WiFi has attracted widespread attention. Prevailing approaches have made great achievements in single location perception and multi-locations fusion perception. However, in practical applications how to realize location-independent sensing using as few samples as possible to achieve high-accuracy recognition is an essential and fairly crucial issue, but still a challenge. To solve the issue, we present a location-independent human activity recognition system based on WiFi named WiLISensing. In this paper, we leverage a simple designed Convolutional Neural Network (CNN) architecture and transfer learning method based on it to recognize activities in a position without training or with very few training samples. What's more, we demonstrate why transfer learning is a better solution to this problem. Extensive experiments have been carried out to show that WiLISensing could achieve promising accuracy above 90% in recognizing six activities and outperform state-of-the-art approaches.