A lightweight deep-learning radar gesture recognition based on a structured pruning-NAS

Abstract

This paper proposes a structured pruning-network architecture search (NAS) algorithm for a lightweight deep-learning radar foot gesture recognition in a conventional lightweight deep-learning models to quantitatively evaluate its performance. Our goal is to recognize foot gestures using a CW radar, generate their STFT unique signatures, and build a foot gesture recognition system that could be implemented on an edge device. The proposed scheme shows that model size and FLOPs were reduced, and a sub-optimal lightweight model for a foot gesture recognition device based on MobileNet was obtained with a slight decrease in accuracy. © 2023 IEEE.