毫米波点云生成论文 阅读笔记 | 3D Point Cloud Generation with Millimeter-Wave Radar

毫米波点云生成论文 | 3D Point Cloud Generation with Millimeter-Wave Radar

Kun Qian, Zhaoyuan He, Xinyu Zhang
UCSD
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (ACM IMWUT)

原始论文地址: http://xyzhang.ucsd.edu/papers/KQian_UbiComp21_RadarPointCloud.pdf
Video地址：ACM SIGCHI官方频道

本文为毫米波点云生成论文 3D Point Cloud Generation with Millimeter-Wave Radar的阅读笔记, 原载于R.X. NLOS的博客
大量参考了作者的 Pre-recorded presentations for UbiComp/ISWC 2021, September 21–26
笔记难免存在问题，欢迎联系 [email protected] 指正。

---

内容在CSDN、知乎和微信公众号同步更新

• CSDN博客
• 知乎
• 微信公众号

---

---

1 Introduction

• Sensors for Automomous Driving
  • Lidar, Camera and Radar
  • Radar is more robust against bad weathers
• Limitations of MmWave Radar (2 main drawbacks)

  • Extremely low resolution

    ❌ due to its small form factor

    ❌ only generates intensity maps with strong reflection peaks

  • Blindness due to specular reflection

    ❌ specularly reflected by most objects

• An Existing Solution: Non-coherent Imaging

  • Fusing measurements along the Radar’s moving trajectory

    ✅ To some extent, alleviate the specular reflection problem (by illuminating from diverse locations)

    ❌ The imaging resolution is still limited by the physical size of the antenna array

    ❌ Cannot be fundamentally improved through spatial sampling

• MilliPoint: Coherent Imaging

  • Raw radar measurements are directly combined with SAR

    ✅ low-end vehicle radars

    ✅ coherently combines measurements of the radar

    ✅ generate dense and high-resolution 3D point clouds

2 困难和解决方法

• Enabling SAR with Vehicle Radar

---

• Challenge 1: SAR requires accurate tracking of the radar
  • 左图：uniform motion (位置已知)
  • 右图: Variable motion (位置未知)、
  • 结论: Without the knowledage of the radar’s location, the image of the object can be highly distorted

• Solution: a self cross-range tracking method

  • Radar Speed = Cross Range Movement / Elapsed Time

  • Observation: Different Tx-Rx pairs experience similar channel responses while moving along the cross-range

    ▪ The radar has equally spaced 4 antennas

    ▪ As the radar moves, the 1st antenna pair at time $t_1$ coincides with the 2nd antenna pair at time $t_0$

    $\Rightarrow$ pair 1和 pair 2接收的信号很接近，只有一个$t_0-t_1$的时延 $\Rightarrow$ 由测量信号可获得 $t_0-t_1$

    ▪ 同时，the radar moves by one antenna spacing $\Rightarrow$ 可获得距离 (cross-range movement)

    ▪ so the speed of the radar can be calculated

---

• Overcoming Specular Reflection in SAR

• Challenge 2: SAR requires focusing on the center of the target

  • 下图：the scenario where a metal surface is 3m away and 沿cross-range方向 135 ${}^{\circ }$ incident angle
  • 左图：SAR focuses on the broadside direction, then the target disappears
  • 右图：only when the SAR focuses on the direction of $45^{\circ }$, the target is imaged

• Solution:

  • root cause of this phenomenon:

    the mismatch of the physical aperture of the radar and the effective apertures of the targets

    The physical aperture: the physical trajectory of the radar (图中blue segment)

    The effective aperture of a target is the segment of the trajectory where the target’s reflection is received $\Rightarrow$ Due to specular reflection, the effective aperture is the projection (图中short red segment)

  • The focusing center can be represented by the parameters $l$ and $\theta$

    ✅ can be identified in the spatial frequency spectrum

    ✅ $\Rightarrow$ 从而propose the multi-focusing algorithm to overcome the specular reflection problem and image all objects in the environment

    ✅ $\Rightarrow$ image all objects in the environment

---

• 3D Point Cloud Generation from SAR
• Challenge 3: SAR only generates 2D intensity maps
  • 下图展示了an indoor scenario
  • images of objects at different heights overlap with each other

• Solution:

  • Combining co-registered pixels of the SAR images generated by different antenna pairs to estimate the height of target scatter point

    ✅ exploit multiple antenna pairs along the vertical direction

    ✅ For each antenna pair, we generate one SAR image and find pixels with high intensity

    ✅ Complex pixel values still encode the elevation information $\Rightarrow$ So we combine the co-registered pixels of all SAR images to form a virtual pixel array and estimate the elevation of the pixels (如下图)

• 点云生成结果

3 实验与结果

• Experiment
  • using a TI cascade radar with 6 transmit antennas and 8 receive antennas
  • Ground Truth: captured by a ZED stereo camera

• Results
  • Compare MilliPoint with non-coherent imaging and the imaging with a static radar
  • In different scenarios with different objects such as cars, pedestrians, bikes and trash bins, MilliPoint generates the densest point cloud s and the points 实现了accurately aligned

总结

• 理论方面 ：Propose MilliPoint

  • The first step to enable 高分辨高密度 3D point cloud generation on low-end vehicle Radar based on SAR

  • 关键挑战：

    Self corss-range tracking

    Automatic multi-focusing

    3D pointcloud generation

• 实验方面 ：implement and verify the design on an automative radar, and conduct case studies in realstic driving scenes

• 更多 ：envision MilliPoint as a new type of sensor fusion modality