Radar-Based Fall Detection Using Multilevel Representations and nMEFMNet

Fall incident is one of the most risky causes of unnatural death of older people, resulting in the urgency of the technology of indoor radar fall monitoring. However, daily activities with similar radar-based characteristics, the motions causing missing and error detection, challenge the mentioned technology. In this article, a millimeter-wave frequency-modulated continuous-wave (FMCW) radar-based fall monitoring technology using multilevel representations and a deep neural network is presented to settle the difficulty. We first propose an approach to obtain three representations from two levels. Specifically, value-filtered time–range (VF-TR) and value-filtered time–velocity (VF-TV) representations in intuitionistic level are used to emphasize the expressible physical property caused by the object, and bispectral representations in nonobjective level are opted to depict indefinable and arcane information, both of which describe the target movement process in their specific level. Their combination could provide the study with a more comprehensive and complementary behavioral representation. Furthermore, we develop a robust, novel, multiextractor, feature-mixed neural network (nMEFMNet) that achieves a superior value of five testing indicators with efficiency. Specifically, nMEFMNet is designed to concatenate the specific tensors in diverse levels captured by the corresponding feature extractors to a single mixed vector, which is crucial to characteristic representation. Moreover, the proposed method achieved a 98.8% classification rate of the whole model and 99.0% accuracy in detecting falls.