Segmentation of Range-Azimuth Maps of FMCW Radars with a Deep Convolutional Neural Network

Abstract

In this paper, we propose a novel deep convolutional neural network for the segmentation of Range-Azimuth (RA) maps produced by a low-power Frequency Modulated Continuous Wave (FMCW) radar to facilitate autonomous operation on a moving platform in a dense, feature-rich environment such as warehouses, storage and industrial sites. Our key contribution is a compact neural network architecture that estimates the extent of free space with high distance and azimuth resolution, yielding high-quality navigation cues at a much lower computational cost than full 2D space segmentation techniques. We demonstrate our method on a unmanned aerial vehicle (UAV) in an industrial warehouse environment.