Low-Power Radar-Based System for Real-Time Object Recognition

Abstract

In recent years, radar technology has attracted a new wave of interest due to unprecedented low-power potential and its inherent low privacy concerns compared to camera systems. In particular, a radar-based object and material recognition encloses a great potential for assistive technology in health-care scenarios, such as prosthetic hands. In this study, we aim to explore such potential via offline and online recognition achieved by deep learning techniques. Twenty different targets are explored, including objects from daily life activity as well as biological tissue (i.e., human hand). Feasibility is confirmed by the offline and online recognition accuracies (achieving 94% and 89% correct classifications in the best case scenario, respectively), and promising insights are offered in regard to the number of radars needed for such a task. Remarkably, for the first time, a radar-based real-time correct differentiation between human tissue and inanimate objects is shown. We believe that these results pave the way for an easy-to-integrate solution with wide potential benefit in industrial automation technology and novel innovative human-machine interfaces, particularly in the prosthetic field context.