Research on Human Falling Recognition Based on Inertial Sensors

Abstract

This article aims to recognize human fall behavior based on wearable inertial sensors. The experiment in this paper mainly adopts data fusion algorithm, which can extract various features that can represent activities in time domain, frequency domain and time-frequency domain from the original data of human motion to effectively distinguish activities. In addition to considering the validity of the data, we also need to consider the way the data relates to the real situation and the comfort of the user in real life. Experimental data shows that in the data collection process, in order to obtain datasets that are easy to calculate, accurate and effective, two aspects need to be considered: the location of the data collection device and the frequency of data collection. Experimental results show that feature extraction has a great influence on the accuracy of activity recognition. Using 6 features of the elderly specifically selected for activity recognition, the original sensor data is directly trained through LSTM-RNN, and the accuracy of activity recognition can reach 92.28%.