Optimum target range bin selection method for time-frequency analysis to detect falls using wideband radar and a lightweight network

• Published in: Biomedical signal processing and control 2022-08, Vol.77, p.103741, Article 103741
• Main Authors: He, Mi, Yang, Yi, Ping, Qinwen, Dai, Ran, Liu, Bingwen, Nian, Yongjian, Zhang, Zhu
• Format: Article
• Language: English
• Subjects: Fall detection; Lightweight network; Time-frequency analysis; Wideband radar
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2022.103741

•The optimum target range bin is defined to obtain the time Doppler spectrogram imagesto distinguish falling from non-falling.•A K-band wideband FMCW radar is applied to build a fall detection database of 36 subjects in two scenes.•The proposed method is superior to the existing maximum variance method.•The test results show that our proposed method is robust enough to detect falls successfully for unseen subjects in unseen situations. Wideband bio-radars have been applied to fall behaviour recognition research in the last few years. The optimum target range bin is that the time Doppler spectrogram image of signals of the range time matrix in this range bin is best for the deep learning network to distinguish between falling and non-falling. Thus, selecting an appropriate target range bin for time-frequency analysis is very important for effective fall detection of wideband radar. A K-band wideband frequency modulated continuous wave radar is applied to build a fall detection database of 36 subjects in two scenes. The radar data are pre-processed to obtain range-Doppler spectrograms where the optimum target range bin is selected for the subsequent time-frequency analysis. The original, threshold, and binary time Doppler spectrograms are compared using different target range bin optimum target range bin selection methods. To determine the effectiveness of the proposed algorithm in fall recognition, a MobileNetV3-Small architecture is implemented for fall detection by automatically extracting features and classifying them. The proposed method distinguishes falls from non-falls with a 98.93% classification accuracy with 5-fold cross validation, which is superior to the existing maximum variance method that is applied to the dataset using three kinds of time Doppler spectrograms. The test results in another scene show that our proposed method is robust enough to detect falls successfully for unseen subjects in unseen situations. The proposed optimum target range bin selection method can effectively and robustly detect falls when using wideband radar and a lightweight network.