Unobtrusive Human Fall Detection System Using mmWave Radar and Data Driven Methods

As the population ages, health issues like injurious falls demand more attention. One solution is to use wearable devices to detect falls. Nevertheless, most of these devices raise obtrusiveness, and older people generally resist or might forget to wear them. The millimeter-wave (mmWave) radar technology was used in this study to unobtrusively detect human falls. Data were collected from healthy young volunteers with the radar mounted on the side wall (trial 1) or overhead (trial 2) of an experimental room. A set of features were manually extracted from the data point clouds; then, multilayer perceptron (MLP), random forest (RF), ${k}$ -nearest neighbor (KNN), and support vector machine (SVM) classifiers were applied on the features. Additionally, we devised a convolutional neural network (CNN)-based deep learning model for the underlying fall detection problem that receives a 3-D representation of the point cloud data, known as occupancy grid, as the input. The optimal installation position of the radar sensor was unknown. Therefore, the sensor was mounted on side wall and on the ceiling of the room to allow the performance comparison between these sensor placements. RF classifier achieved the best results in trial 2 (an accuracy of 92.2%, a recall of 0.881, a precision of 0.805, and an ${F}1$ -score of 0.841), and the proposed CNN model achieved slightly better results comparing to the RF method in trial 2 (an accuracy of 92.3%, a recall of 0.891, a precision of 0.801, and an ${F}1$ -score of 0.844). These results suggest that the development of an unobtrusive monitoring system for fall detection using mmWave radar is feasible.