DenserRadar: 4D millimeter wave radar point cloud detector based on dense LiDAR point cloud

Original title: DenserRadar: A 4D millimeter-wave radar point cloud detector based on dense LiDAR point clouds

Paper link: https://arxiv.org/pdf/2405.05131

Author's unit: Tsinghua University

Thesis idea:

4D millimeter wave (mmWave) radar is used in extreme environments Its robustness, wide detection range, and ability to measure speed and altitude have shown significant potential to enhance perception when autonomous driving systems face corner-cases. However, the inherent sparsity and noise of 4D millimeter-wave radar point clouds limit its further development and practical applications. This paper introduces a novel 4D millimeter-wave radar point cloud detector that exploits high-resolution dense radar point clouds. Our method constructs dense 3D occupied space ground truth from stitched LiDAR point clouds and uses a specially designed network called DenserRadar. The proposed method surpasses existing probability-based and learning-based millimeter-wave radar point cloud detectors in terms of point cloud density and accuracy, achieving better results on the K-Radar dataset.

Main contribution:

The work of this paper is the first 4D millimeter wave radar point cloud detector supervised by dense 3D occupied data space ground truth. Generated by splicing multi-frame LiDAR point clouds, thereby densifying the detected millimeter wave radar point cloud.

This paper proposes an innovative dense snow 3D occupied data space ground truth generation process, and the splicing of dense LiDAR point clouds of the K-Radar dataset. These point clouds provide comprehensive scene ground truth. In the publication will be available for further research.

Due to the special design of the DenserRadar network, the algorithm in this article is superior to existing CFAR type and learning-based millimeter wave radar point cloud detection methods in terms of point cloud density and accuracy.

Network Design:

Autonomous driving technology aims to provide a safe, convenient and comfortable transportation experience, and its development speed is impressive. To achieve high-level autonomous driving, the ability to perceive and position complex environments is indispensable. Therefore, the sensors equipped on autonomous vehicles, including cameras, lidar (LiDAR) and millimeter-wave radar, as well as their related algorithms, are attracting more and more research interest.

Given its advantages such as compact size, high cost-effectiveness, all-weather adaptability, speed measurement capability, and wide detection range [1], millimeter wave (mmWave) radar has been widely used in the field of autonomous driving. Recent advances in multiple-input multiple-output (MIMO) antenna technology have further improved height resolution, enabling the emergence of 4D millimeter-wave radar. Therefore, 4D millimeter-wave radar is increasingly regarded as a key enhancement of perception and positioning capabilities in autonomous driving, especially in challenging edge scenes such as rain, snow, and fog. As its name suggests, 4D millimeter wave radar can measure four dimensions of target information: range, azimuth, altitude and Doppler velocity, providing a comprehensive sensing solution.

However, the quality of 4D millimeter wave radar point clouds lags significantly behind lidar point clouds. First of all, 4D millimeter wave radar point clouds have low resolution problems, especially in angle measurement. This limitation is mainly due to the antenna configuration and direction of arrival (DOA) estimation of mmWave radar [2]. Secondly, 4D millimeter wave radar point clouds are much sparser than lidar point clouds. Third, due to multipath effects, signal interference and ground reflection, 4D millimeter wave radar point clouds often contain a large number of clutter points. All these shortcomings hinder the application of 4D millimeter wave radar in autonomous driving.

The quality of 4D millimeter wave radar point clouds is not only limited by hardware, but also by signal processing algorithms [3]. In particular, detecting real targets from raw radargrams or tensors to generate point clouds can greatly impact quality. Traditionally, the False Alarm Rate (CFAR) detector and its variants [4], [5] are widely used in the detection of millimeter wave radar point clouds. However, as a probability-based algorithm, CFAR-type detectors may encounter problems when detecting objects of different sizes because these objects are not independent and identically distributed [6], which often occurs in autonomous driving scenarios.

In order to solve the point cloud quality issues related to 4D millimeter wave radar, this paper proposes a learning-based 4D millimeter wave radar point cloud detector, which is composed of dense real-life data generated from lidar point clouds. Information supervision. Initially, this paper stitches multiple frames of pre-processed LiDAR point clouds to generate dense 3D occupancy ground truth. This article then introduces the DenserRadar network, which extracts the features of the original 4D millimeter wave radar tensor and generates a denser and more accurate 4D millimeter wave radar point cloud. The network employs a weighted hybrid loss function along with other novel design elements to capture multi-resolution features and generate point clouds with better resolution than traditional techniques. Comparative experiments conducted on the K-Radar data set [7] prove the effectiveness of this method.

The algorithm of this article is shown in Figure 1. First, this paper designs a ground truth generation process to obtain dense 3D occupied space ground truth as supervision information by splicing multi-frame lidar point cloud data, and then establishes the DenserRadar network, which is tasked with generating data from the original 4D millimeter wave Detection of millimeter wave radar point clouds in radar tensor data.

Figure 1. Overview of the entire algorithm.

Figure 2. Truth value generation flow chart.

Experimental results:

Figure 4. Qualitative point cloud comparison between the DenserRadar algorithm and the CA-CFAR algorithm in this article, attached Images and dense 3D space-occupying ground truth point clouds are used as reference. Each arrow in the diagram represents a length of 10 meters.

Summary:

This article introduces DenserRadar, a novel 4D millimeter wave radar point Cloud detection networks, and an innovative process for generating dense ground truth. Experimental results and ablation studies demonstrate the effectiveness of our network architecture and ground truth generation methodology. This research has the potential to improve the perception and localization capabilities of autonomous driving systems, especially in challenging edge-case scenarios.