Synthesis of Micro-Doppler Signatures for Abnormal Gait using Multi-branch Discriminator with Embedded Kinematics

A key limiting factor in the depth, hence accuracy of deep neural networks (DNNs) designed for radar applications, is the meager amount of data typically available for training. Generative adversarial networks (GANs) have been proposed in many fields for the generation of synthetic data. It was shown, however, that when applied to micro-Doppler signature simulation, GANs suffer from performance degradation due to the generation of kinematically impossible samples. In this work, kinematic analysis of the micro-Doppler signature envelope is integrated as an additional branch in the discriminator network of a GAN to improve the kinematic fidelity of synthetic data when simulating abnormal gait signatures. Results show that the proposed multi-branch GAN network results in greater overlap in the feature space of synthetic abnormal gait samples with that of measured signatures for abnormal gait.