Title: A MIMO Radar-Based Metric Learning Approach for Activity Recognition

Abstract: Human activity recognition is seen of great importance in the medical and surveillance fields. Radar has shown great feasibility for this field based on the captured micro-Doppler ({\mu}-D) signatures. In this paper, a MIMO radar is used to formulate a novel micro-motion spectrogram for the angular velocity ({\mu}-{\omega}) in non-tangential scenarios. Combining both the {\mu}-D and the {\mu}-{\omega} signatures have shown better performance. Classification accuracy of 88.9% was achieved based on a metric learning approach. The experimental setup was designed to capture micro-motion signatures on different aspect angles and line of sight (LOS). The utilized training dataset was of smaller size compared to the state-of-the-art techniques, where eight activities were captured. A few-shot learning approach is used to adapt the pre-trained model for fall detection. The final model has shown a classification accuracy of 86.42% for ten activities.