A Multi-class Dataset Expansion Method for Wi-Fi-Based Fall Detection

Nowadays, with the wide commercial use of Wi-Fi technology, the use of Wi-Fi channel state information (CSI) for fall detection has gradually become a hot research field. However, many existing fall detection systems based on Wi-Fi lack accurate action classification because of the high acquisition cost of complex action datasets. They cannot accurately identify complex fall actions, and have a high false positive rate. This paper proposes a multi-class dataset expansion method for different fall actions and non-fall actions, which classifies the movements in detail according to fall speed and other limb movements and expands the scale of the data set by dividing and reorganizing the limited data. As a result, the proposed method reaches a recognition accuracy of 91.6%.

I. Introduction

According to statistics, fall accidents have become the second leading cause of unintentional injuries[1]. Most fall-related deaths occur in people over the age of 65, so fall has become one of the important factors endangering the elderly. Facing the increasing global aging problem, developing a set of portable, accurate, and real-time fall detection system has extremely important economic and social value. In fact, over the past few decades, people have been creating new technologies to realize more effective human behavior perception and analysis. The traditional technical methods mainly include the following ways: computer vision [5], infrared technology [6] and special sensor technology [7]. However, most of these methods need to deploy expensive additional devices, and there may be a risk of divulging customer privacy. Some traditional detection methods require wearables and are thus not user-friendly.