MAROAM Map-based Radar SLAM through Two-step Feature Selection Dequan Wang Yifan Duan Xiaoran Fan Chengzhen Meng Jianmin Ji and Yanyong Zhang

2025-05-02 0 0 1.4MB 7 页 10玖币

侵权投诉

MAROAM: Map-based Radar SLAM through Two-step Feature

Selection

Dequan Wang, Yifan Duan, Xiaoran Fan, Chengzhen Meng, Jianmin Ji and Yanyong Zhang

Abstract— In this letter, we propose MAROAM, a millimeter

wave radar-based SLAM framework, which employs a two-

step feature selection process to build the global consistent map.

Speciﬁcally, we ﬁrst extract feature points from raw data based

on their local geometric properties to ﬁlter out those points

that violate the principle of millimeter-wave radar imaging.

Then, we further employ another round of probabilistic feature

selection by examining how often and how recent the feature

point has been detected in the proceeding frames. With such

a two-step feature selection, we establish a global consistent

map for accurate and robust pose estimation as well as other

downstream tasks. At last, we perform loop closure and graph

optimization in the back-end, further reducing the accumulated

drift error.

We evaluate the performance of MAROAM on the three

datasets: the Oxford Radar RobotCar Dataset, the MulRan

Dataset and the Boreas Dataset. We consider a variety of

experimental settings with different scenery, weather, and road

conditions. The experimental results show that the accuracy of

MAROAM is 7.95%, 37.0% and 8.9% higher than the currently

best-performing algorithms on these three datasets, respectively.

The ablation results also show that our map-based odometry

performs 28.6% better than the commonly used scan-to-frames

method. Finally, as devoted contributors to the open-source

community, we will open source the algorithm after the paper

is accepted.

I. INTRODUCTION

In recent years, millimeter wave radar (radar, in short) has

received an increasing amount of attention in autonomous

systems. Compared to vision and LiDAR, radar offers longer

detection ranges and more robust performance under adverse

lighting and weather conditions, lending it towards a viable

sensing option for applications like self driving and robotics

[1]. However, in order for radar to grow from a supple-

mentary sensing approach to one of the primary ones, their

sensing capability, especially when used alone, needs to be

carefully investigated and strengthened. Towards this goal,

we study how to utilize the radar alone for the Simultaneous

Localization and Mapping (SLAM) task in this work.

The main weakness for millimeter wave radar images

lies in the fact that they suffer from the multipath effect,

sidelobes, clutter, etc. [1], which may lead to false alarms. In

order to overcome the effect of false alarms, the recent radar

SLAM systems mainly focus on extracting robust features

from radar images. For example, Cen et al. proposed to

extract features using traditional signal processing techniques

School of Computer Science and Technology, University

of Science and Technology of China, Hefei, 230026, China

{wdq15588, dyf0202, czmeng}@mail.ustc.edu.cn,

ox5bc@winlab.rutgers.edu, {jianmin,

yanyongz}@ustc.edu.cn.

Fig. 1. An example radar point cloud map with global consistency, which

is built by MAROAM with the Oxford Radar RobotCar Dataset. For clarity,

we also zoom in several locations and show the detailed views. In this

example, we use the ”17-13-26” sequence with a total length of 10km.

[2], [3], while Barnes et al. proposed end-to-end learning-

based feature extraction methods [4], [5]. However, even with

these efforts, feature extraction for radar systems remains

an active research topic, which aims at minimizing the

cumulative errors in the feature matching task.

Next, let us take a close look at the radar odometry. When

the scan-to-map registration strategy is engaged, the number

of features grows rapidly with the map, a large portion

of which are of low quality and should not be selected.

The quality of the map seriously affects the performance

of map based-SLAM [6]. RadarSLAM [7] uses descriptor-

based pairwise matching of features. However, since image

noise and different radar cross-sectional (RCS) generated by

different viewing angles [8] will directly affect the repre-

sentation of descriptors, so it has high requirements for the

extraction process of features and descriptors.

Since pairwise matching of descriptors is not required,

the Iterative Closest Point (ICP) [9] methods achieve state-

of-the-art accuracy in motion estimation, such as CFEAR

RadarOdometry [10] and NDT [11], but the submaps they

use are only the superposition of multiple frames, and there

is no mapping process. The disadvantage of ICP-based point

cloud registration algorithms is that they are sensitive to

initial values and tend to converge to local optimum [12].

arXiv:2210.13797v1 [cs.RO] 25 Oct 2022

Since radar point cloud has more noise, the ICP-based algo-

rithm is more likely to fall into a local optimum, resulting

in a larger motion estimation error. We have demonstrated

this experimentally at IV-C. Considering the importance of

map quality, our paper looks at how feature selection can be

used to optimize the process of map building and scan-to-

map matching in ICP-based radar SLAM, resulting in more

accurate maps and more robust state estimates.

In order to address these limitations, we propose a Map-

based Radar Odometry and Mapping system, MAROAM,

which is an accurate and robust map-based radar SLAM

framework based on LOAM [13]. In order to improve

map quality, we use a two-step geometry-probability fea-

ture selection strategy in the registration and map update

stage. In speciﬁcity, we ﬁrst extract surface features by

calculating the Local Linearity and the Local Aggregation

with a geometry-based ﬁlter. We then involve a point-to-line

ICP method, using the surface features to solve the relative

pose transformation. Subsequently, we use a probability-

based ﬁlter to dynamically ﬁlter out the features that are

not frequently selected in the ICP matching process. Using

the more frequent features, our algorithm realizes the scan-

to-map matching strategy of ICP-based radar SLAM for the

ﬁrst time. Additionally, We use the Scan Context [14] for

loop closure and modify it to ﬁt the radar data format.

We evaluate the performance of MAROAM on the three

datasets: the Oxford Radar RobotCar Dataset [7], the Mul-

Ran Dataset [15] and the Boreas Dataset [16]. We consider

a variety of experimental settings with different scenery,

weather, and road conditions. The experimental results show

that the accuracy of MAROAM is 7.95%, 37.0% and 8.9%

higher than the currently best-performing algorithms on these

three datasets, respectively. The ablation results also show

that our map-based odometry performs 28.6% better than

the commonly used scan-to-frames method.

In summary, our main contributions are as follows:

•We propose MAROAM, an map-based radar SLAM

framework based on LOAM. Using a geometry-

probability two-step feature selection, our algorithm

realizes the scan-to-map matching strategy of ICP-based

radar SALM for the ﬁrst time.

•We evaluate our MAROAM on three datasets which

contain a variety of scenarios, weathers, and road condi-

tions. Experiments show that our algorithm outperforms

SOTA algorithm on all three datasets.

II. RELATED WORKS

A. Radar SLAM

We divide the existing radar odometry/SLAM algorithms

into four categories according to the method of estimating

the relative pose:

1) Direct methods: Checchin et al. [17] ﬁrst time use the

Fourier-Mellin Transform (FMT) to register radar images in

a sequence for motion estimation. PhaRao [18] apply FMT

to Cartesian and log polar radar images to estimate rotation

and translation to decouple rotation and translation.

2) ICP-based methods: Adolfsson et al. [10] process the

landmarks extracted by CFAR through k-strongest ﬁltering

in Cartesian coordinates, and optimizes the point-to-line

optimize metric to estimate the relative pose. Kung et al.

[19] propose a RO with probabilistic submap building, and

an NDT-based radar scan matching. Both of them use scan-

to-submap matching, but the submaps they use are only the

superposition of multiple frames, and there is no mapping

process.

3) Descriptor-based methods: Considering the unneces-

sary inﬂuence of radar echo, Cen et al. [2] propose an

algorithm to extract landmarks, and performe scan match-

ing by greedily adding features correspondence based on

unary descriptor and pairwise compatibility score. After

that, another feature extraction algorithm that only use one

parameter is proposed by Cen et al. [3], and graph matching

is used for scan matching. Hone et al. [20], [21] propose

a full radar-based SLAM pipeline, RadarSLAM, composed

of pose tracking, local mapping, loop closure detection and

pose graph optimization. RadarSLAM uses visual features

for scan matching and M2DP [22] descriptor based on point

cloud for loop detection.

4) Learning-based methods: Barnes et al. [5] use deep

neural network to learn an embedding space that is basically

free of artifacts and interference, which is used to perform ef-

fective correlation matching between continuous radar scans,

and achieve high accuracy without considering spatial cross-

validation(CSV). A self supervised learning framework is

proposed by Barnes et al. [4] to detect the robust key features

of range estimation and metric positioning in radar. Burnett et

al. [23] uses unsupervised method to extract features, which

increases generalization and performs well on both Oxford

and Boreas.

B. LiDAR SLAM

LOAM [13] is the pioneering work of 3D LiDAR SLAM,

which proposes a basic framework. LOAM extracts edge

features and plane features from the original point cloud,

and designs related loss functions for each type of feature.

The matching consists of a fast frame-to-frame match and a

slow frame-to-graph match. F-LOAM [24] follows LOAM

and abandons the scan-to-scan match and replaces it by

only scan-to-map with high frequency. Duan et al. [25]

propose a feature ﬁlter, pFilter, by properly measuring each

feature point’s p-Index and only keeping those with high

index values, and improved both the efﬁciency and accuracy

of the registration process. Our work is inspired by the

aforementioned LiDAR SLAM.

III. METHOD

A. System Overview

An overview of MAROAM is shown in Fig. 2. The system

receives input from a 2D radar image and outputs a 3-

DOF pose estimation and a global consistent map. The

SLAM system consists of two modules: tracking and loop

closure. The tracking module is used for feature extraction,

feature ﬁltering and real-time motion estimation, and the

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

MAROAM:Map-basedRadarSLAMthroughTwo-stepFeatureSelectionDequanWang,YifanDuan,XiaoranFan,ChengzhenMeng,JianminJiandYanyongZhangAbstractInthisletter,weproposeMAROAM,amillimeterwaveradar-basedSLAMframework,whichemploysatwo-stepfeatureselectionprocesstobuildtheglobalconsistentmap.Specically,werstextr...

展开>> 收起<<

MAROAM Map-based Radar SLAM through Two-step Feature Selection Dequan Wang Yifan Duan Xiaoran Fan Chengzhen Meng Jianmin Ji and Yanyong Zhang.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

MAROAM Map-based Radar SLAM through Two-step Feature Selection Dequan Wang Yifan Duan Xiaoran Fan Chengzhen Meng Jianmin Ji and Yanyong Zhang

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: