Deep Learning Wi-Fi Channel State Information for Fall Detection

It is common that the elderly may fall and injure severely. This problem has attracted worldwide attention and becomes a major challenge in the public health care. In the past decade, extensive studies have been conducted to detect fall using wearable sensors and cameras. Given the pervasive WiFi penetration in our daily life, behavior recognition based on the channel state information (CSI) of WiFi signals has shown its potentials in detecting falls for the elderly with less constraint compared with clumsy sensors. In this paper, we conducted a performance evaluation study of three deep learning methods on a public dataset to detect falls. The experiment results show that the accuracy of the deep learning algorithms on Wi-Fi datasets achieves beyond 95% which may generate notable market values. Nevertheless, the long training time of deep learning models is likely to be the hampering factor before commercialization. Our study may stimulate further research on accelerating deep learning methods in a software/hardware co-design approach.