Fall Detection Based on Deep Learning Fusing Ultrawideband Radar Spectrograms

Abstract

Compared with narrowband Doppler radar, ultrawideband radar can simultaneously acquire the range and Doppler information of targets, which is more beneficial for behavior recognition. To improve the recognition performance of fall behavior, frequency-modulated continuous-wave ultrawideband (UWB) radar was applied to collect daily behavior and fall data of 36 subjects in two real indoor complex scenes, and a multi-scene fall detection dataset was established with various action types; the range-time, time-Doppler, and range-Doppler spectrograms of the subjects were obtained after preprocessing radar data; based on the MobileNet-V3 lightweight network, three types of deep learning fusion networks at the data level, feature level, and decision level were designed for the radar spectrograms, respectively. A statistical analysis shows that the decision level fusion method proposed in this paper can improve fall detection performance compared with those using one type of spectrogram, the data level and the feature level fusion methods (all P values by significance test method are less than 0.003). The accuracies of 5-fold cross-validation and testing in the new scene of the decision level fusion method are 0.9956 and 0.9778, respectively, which indicates the good generalization ability of the proposed method.

Keywords

• ultrawideband radar
• deep learning
• fall detection
• data fusion
• ligthweight network