ADCNet: Learning from Raw Radar Data via Distillation

TL;DR

This paper introduces ADCNet, a novel radar perception model that leverages raw radar data and signal processing techniques to improve detection accuracy for autonomous driving applications.

Contribution

ADCNet integrates signal processing into neural networks and uses pre-training on pseudo-labeled data to enhance radar perception performance.

Findings

Achieves state-of-the-art detection on RADIal dataset

Effectively distills signal processing information into a fast neural network

Demonstrates robustness in adverse weather conditions

Abstract

As autonomous vehicles and advanced driving assistance systems have entered wider deployment, there is an increased interest in building robust perception systems using radars. Radar-based systems are lower cost and more robust to adverse weather conditions than their LiDAR-based counterparts; however the point clouds produced are typically noisy and sparse by comparison. In order to combat these challenges, recent research has focused on consuming the raw radar data, instead of the final radar point cloud. We build on this line of work and demonstrate that by bringing elements of the signal processing pipeline into our network and then pre-training on the signal processing task, we are able to achieve state of the art detection performance on the RADIal dataset. Our method uses expensive offline signal processing algorithms to pseudo-label data and trains a network to distill this information into a fast convolutional backbone, which can then be finetuned for perception tasks. Extensive experiment results corroborate the effectiveness of the proposed techniques.