Human Interpretable Radar Through Deep Generative Models

Abstract

Imaging radars are a powerful new sensor aimed at the autonomous vehicle market. With high angular resolution and dynamic range, such radars can discern between different adjacent objects, even when those are stationary. The output of the radar is a complex Point Cloud (PC) that is difficult for humans to understand and recognize objects. We investigate the use of machine learning to transform the radar PC to an interpretable format that can be understood intuitively. We employ two different generative models to transform radar PCs into synthetic LiDAR PCs and camera images of the scene. Our results demonstrate the imaging radar's ability to recognize objects such as pedestrians, parked vehicles, trees and road edges. We show how these tools can be used to analyse and evaluate the radar PC.