FMCW Radar-Based Human Activity Recognition Based on Higher Order Synchrosqueezing Transform

Human activity recognition (HAR) plays a fundamental role in various health care and smart home automation. The occurrence of falls is a key focus in HAR, due to its implications in safety-critical applications, such as elderly monitoring and assisted living. In previous studies, radar-based HAR utilizes discrete Fourier transform (DFT), short-time Fourier transform (STFT), and wavelet transform-based approaches, but they fail due to poor resolutions and fixed basis function to distinguish closely spaced frequency components in the radar signals. To mitigate these limitations, we present a new method for HAR based on the higher order synchrosqueezing transform and deep learning (DL) classifiers from radar return signals. The frequency modulated continuous wave (FMCW) radar return signal captures the micro- and macro-motion of human activity. To analyze such signals, the fourth-order Fourier synchrosqueezing transform (FSST4) technique plays a vital role due to its impressive time–frequency resolutions in the resultant time–frequency representation (TFR). The DL techniques (MobileNetV2, GoogLeNet, AlexNet, VGG16, and VGG19) are used to classify the FSST4-based TFRs into various activity classes. The method based on FSST4 and AlexNet achieved the highest accuracy of 99.40% among the studied classifiers. A comparative study is also performed to study the effectiveness of FSST4-based TFRs as compared with STFT, continuous wavelet transform (CWT), Fourier synchrosqueezing transform (FSST), second-order Fourier synchrosqueezing transform (FSST2), third-order Fourier synchrosqueezing transform (FSST3), and FSST4-based TFRs in the proposed method. The proposed method with FSST4-based TFR provided superior performance as compared with other TFR-based proposed methods. The comparison of computational complexity analysis of FSST4 with other methods, such as STFT, CWT, FSST, FSST2, and FSST3, is also studied in this work. The CWT is found to be the fastest among studied methods for TFR computation. The FSST4 method required a similar computation time compared with STFT, FSST, FSST2, and FSST3 for obtaining TFRs providing relatively better performance.