[论文阅读] DAY 1 Vehicle Detection With Automotive Radar Using Deep Learning on Range-Azimuth-Doppler

Employing deep learning techniques, the detection of vehicles is facilitated through automotive radar operating on range-azimuth-doppler tensors.

作者: Qualcomm AI Research∗ Qualcomm Technologies, Inc.†

该研究主要关注基于自动雷达的显著车辆检测技术及其在范围-角度-多普勒域中的深度学习应用。本文提出了一种基于自动雷达实现高效车辆检测的新方法，并将其与深度学习模型相结合以提高检测性能。通过引入先进的特征提取算法和优化策略，在有限计算资源下实现了系统的高效运行。实验结果表明，在相同的硬件条件下，该系统能够达到或超越现有先进算法的性能水平。

Abstract

二十多年来一直是汽车高级驾驶员辅助系统的可靠技术核心。

Radar has long served as a leading technology in advancing driver-assistance systems within the automotive industry for over two decades.
As an affordable, all-weather sensor with extensive range capabilities, radar is projected to play an indispensable role in shaping autonomous driving technologies. Conventional radar signal processing methods often struggle to differentiate between reflections from objects of interest and background clutter, typically confined to detecting signal peaks.
These peak-based techniques effectively transform the image-like nature of radar signals into sparse point clouds. In this research, we present a novel deep-learning-based vehicle detection solution that processes image-like tensors directly rather than relying on point clouds derived from peak detection.
To our knowledge, this marks the first implementation of such an approach.

论点

• Radar 可被视为实现低成本解决方案的有效手段之一。
  • 在对光线变化、雨天及雾霾天气等方面的稳定性上，Radar 明显优于相机（Camera）和Lidar。
  • 由于雷达（Radar）通过测量电磁反射强度来识别障碍物，在直观表示这一过程时仍面临诸多挑战。
  • 传统的数据处理方式将原始信息（包括三维张量和由多普勒效应生成的速度值）转化为稀疏二维点云图谱，并将障碍物与其环境进行分离；这种做法会导致大量关键信息丢失。
  • 数据处理利用快速傅里叶变换（FFT），原始数据被转换为频域中的三维张量，
  • 并随后通过深度学习模型进行分析。

结果