Synthetic Radar Signal Generator for Human Motion Analysis

Abstract

Synthetic generation of radar signals is an attractive solution to alleviate the lack of standardized datasets containing paired radar and human-motion data. Unfortunately, current approaches in the literature, such as SimHumalator, fail to closely resemble real measurements and thus cannot be used alone in data-driven applications that rely on large training sets. Consequently, we propose an empirical signal model that considers the human body as an ensemble of extended targets. Unlike SimHumalator, which uses a single-point scatterer, our approach locates a multiple-point scatterer on each body part. Our method does not rely on 3-D-meshes but leverages primitive shapes fit to each body part, thereby making it possible to take advantage of publicly available motion-capture (MoCap) datasets. By carefully selecting the parameters of the proposed empirical model, we can generate Doppler-time spectrograms (DTSs) that better resemble real measurements, thus reducing the gap between synthetic and real data. Finally, we show the applicability of our approach in two different application use cases that leverage artificial neural networks (ANNs) to address activity classification and skeleton-joint velocity estimation.