FilterNet: A Convolutional Neural Network for Radar-Based Fall Detection by Filtering Out Non-fall Feature in the Spectrogram

Fall detection is an emerging topic for health care and smart surveillance. Non-intrusive radar-based fall detection is a challenging but promising problem. Compared with ideal fall samples in the training set, the fall events in practical situations happen unintentionally and usually contain additional features in the time-frequency spectrogram because the radar senses other moving events. We defined this problem as "feature doping" of the radar spectrogram. In this work, we proposed a novel training strategy to train a CNN-based filter named FilterNet to tackle the feature doping problem in time-frequency spectrogram analysis for fall detection, which will severely deteriorate the performance of existing binary classification methods. Our method trains the FilterNet to fully reconstruct the radar spectrogram of fall motions while generating zero images for non-fall motions' spectrograms. Through the proposed FilterNet, the features of fall remain, and those of non-fall are eliminated in spectrograms in the filtered domain so that the subsequent classifier can get rid of the feature doping problem and detect fall events easily. Experimental results show that our method outperforms the state-of-the-art methods on the testing dataset containing fall motions with additional interferential features and prove the effectiveness of our method in handling the practical problem of fall detection using spectrogram.