Synchro-Squeezed Time-Frequency Representations for Radar-based Human Activity Recognition

Time-frequency representations such as short-time Fourier transform and wavelet transform of the radar returns capture macro and micro motions of the individuals that facilitate human activity recognition using radars. However, the choice of windows in these transforms constrains the time-frequency resolution and may negatively impact the recognition of human activities. Synchro-squeezed Fourier and synchrosqueezed wavelet transforms are known to provide improved time-frequency resolution and therefore may provide better human activity recognition. This work uses the histogram of gradients (HOG) of the synchro-squeezed transforms as hand-crafted features for radar-based human activity recognition. Two types of synchro-squeezed time-frequency representations, namely, synchro-squeezed Fourier transform (SSFT) and synchro-squeezed wavelet transform (SSWT) are considered in this work. HOG features obtained from the synchrosqueezed transforms, the traditional short-time Fourier transform (STFT), and continuous time wavelet transform (CWT) are used with four well-known machine-learning classifiers, namely stochastic gradient descent (SGD), random forest (RF), K-nearest neighbor (KNN) and support vector machine (SVM). A publicly available dataset consisting of radar signatures of human activities recorded at three different locations is used in this study and a location-wise training/testing strategy is utilized. The performance of radar-based activity recognition is significantly improved with the use of synchro-squeezed time-frequency representations as compared to time-frequency representations without synchro-squeezing.