Fall Detection Using Probability Density Based Denoising Algorithm and 3D Doppler-Time Maps

With the increasing population of the elderly, their health problems have attracted much concern. In particular, falls are a common cause of disability and even death among older adults, and thus, fall monitoring is becoming a necessity for them. However, for the fall detection algorithms based on the Doppler time (DT) map, the factors, such as the noise and clutter, will significantly affect the detection accuracy. Therefore, we propose a denoising algorithm based on Rayleigh probability distribution for DT maps. To fuse micro-Doppler features of human movements, we then construct them as a three-dimensional (3D) matrix with the vectors near the target range bin. Rather than extract 2D features in DT maps, we put the 3D-DT maps into a lightweight neural network composed of 3D convolutional neural networks (3D-CNNs) bi-directional long short term memory (Bi-LSTM) networks designed in this paper In order to test the effect of the denoising algorithm, we collect four daily motions and three fall motions. And we find the performance of the network was significantly improved compared with mean filtering.