Lightweight Fall Detection Based on FMCW Micro-Doppler and Power Burst Signatures Extraction

Nowadays, the aging problem is intensified, and the demand for the elderly fall detection at home is increased. The fall detection system based on millimeter-wave frequency modulated continuous wave (FMCW) radar shows its advantages. This paper proposes a lightweight fall detection system based on FMCW radar. The noise is removed by ordered statistics constant false alarm rate (OS-CFAR) filter. After that, the power burst curve (PBC) is used to detect motion events. The micro-Doppler frequency estimations of human body parts are extracted by CLEAN, density-based spatial clustering of applications with noise (DBSCAN) and k-nearest neighbors (KNN) algorithm. Finally, 1D convolutional neural network (1D-CNN) and bidirectional long short-term memory (Bi-LSTM) network are used to make the final detection. The accuracy of the fall detection model is 93.23%, with precision, recall and F1 score exceeding 90%.