Wifi-Based Human Fall Detection Using Machine Learning


Gayatri Pise

Oral Defence Date: 

Wednesday, July 24, 2019 - 11:00


TH 434


Prof. Hao Yue and Prof. James Wong


According to a recent report from the World Health Organization, human falls are the leading cause of accidental deaths worldwide among older adults. A fall is so common that every 17 seconds an older person is treated for fall related injuries, and every 30 minutes an older adult dies due to fall complications. Within 30 to 60 minutes of falling, muscle cells of the human body begin to breakdown, which potentially results in dehydration, pressure sores, hypothermia and pneumonia. This damage can be reduced if there is a rapid response and rescue of the falling person. Currently, there are fall detection mechanisms that alerts emergency services when they identify a human fall event. However, they are either battery dependent and must be worn by the person all the time or they are camera based and invade privacy of the person by monitoring continuously. Recent wireless studies indicate that as the Wi-Fi signals propagate through an environment, they collect valuable information about their propagation environment. This information is sufficient to detect human activity recognition. We take the learnings from such studies and use it to develop a system that can detect human fall events autonomously, accurately and in real time. We achieve this by extracting the Channel State Information (CSI) from Wi-Fi signals in environments with and without human fall events. We then construct and train a machine learning classifier consisting of an Auto-encoder and SoftMax regression with the CSI data to recognize human fall. The classifier is trained with diverse datasets in different environment settings and can achieve an accuracy of up-to 92% to detect human fall events. These results show great promise that wireless sensing technologies can be used for detecting human falls and alert the emergency medical services to provide necessary help.

Gayatri Pise

Computer Networks, Wi-Fi Signals, Channel State Information, Machine Learning, Auto-encoder, SoftMax regression