激光感知-地面分割-patchwork论文

preview
试读
8页
patchwork
需积分: 0 2 下载量 91 浏览量 更新于2023-12-08 1 收藏 5.52MB PDF 举报
《激光感知：基于同心圆区的区域化地面分割——Patchwork方法》 在现代移动平台，如无人地面车辆(UGVs)、无人机(UAVs)和自动驾驶汽车中，对周围环境的感知需求日益增长。为了实现这一目标，众多研究者已经采用各种三维感知方法。其中，三维激光雷达(LiDAR)传感器因其厘米级的精度、全方位的感知能力以及与立体相机相比更远的测量距离而被广泛使用。然而，地面分割是这些平台进行导航或邻近物体识别的关键，而现有的地面分割算法在处理不平整地面时面临挑战，例如陡峭的坡度、凹凸不平的道路以及路缘石、花坛等障碍物。 针对这些问题，本文提出了一种名为Patchwork的新颖地面分割方法，它对下分割问题有较强的鲁棒性，并且运行频率超过40Hz。Patchwork方法的核心在于将点云编码为基于同心圆区模型的表示，以在各个区间内分配适当密度的云点，同时保持较低的计算复杂度。随后进行区域化地面平面拟合，以估计每个区间的局部地面。引入地面概率估计，显著减少误检。通过在SemanticKITTI和粗糙地形数据集上的实验验证，提出的Patchwork方法相比于最先进的方法展现出颇具潜力的性能，而且在速度上优于现有的平面拟合方法。 具体来说，文章首先介绍了一种将点云编码到同心圆区模型的方法。这种方法将空间划分为多个同心圆环，每个环代表一个特定的距离范围。通过这种方式，可以更好地捕捉地面的不规则形状，同时降低了计算资源的需求。接着，区域化地面平面拟合阶段，每个圆环内的点被用来独立估计局部地面平面，这样能够处理复杂的地形情况，如坡地和小凸起。这种方法比全局平面拟合更能精确地捕捉地面的变化。 然后，地面概率估计是解决误检问题的关键。通过对每个点的地面概率进行估计，可以更准确地区分地面点和其他非地面点，从而减少错误分类。这一步骤结合了之前步骤的成果，提高了分割结果的准确性。 实验部分，文章对比了Patchwork方法在SemanticKITTI和粗糙地形数据集上的表现，证明了其在处理复杂环境下的优越性能。此外，代码的开源性（可在https://github.com/LimHyungTae/patchwork获取）使得其他研究者和开发者能够进一步测试和改进这个方法。 Patchwork方法通过创新的点云处理策略和地面分割技术，提升了移动平台在不平整地面上的导航和物体识别能力。这种方法不仅在性能上有所突破，而且在实时性方面也具有优势，为未来智能系统的地面感知提供了一条新的研究路径。
1
Patchwork: Concentric Zone-based Region-wise
Ground Segmentation with Ground Likelihood
Estimation Using a 3D LiDAR Sensor
Hyungtae Lim
1
, Student Member, IEEE, Minho Oh
1
, Hyun Myung
1
, Senior Member, IEEE
Abstract—Ground segmentation is crucial for terrestrial mo-
bile platforms to perform navigation or neighboring object
recognition. Unfortunately, the ground is not flat, as it features
steep slopes; bumpy roads; or objects, such as curbs, flower beds,
and so forth. To tackle the problem, this paper presents a novel
ground segmentation method called Patchwork, which is robust
for addressing the under-segmentation problem and operates
at more than 40 Hz. In this paper, a point cloud is encoded
into a Concentric Zone Model–based representation to assign an
appropriate density of cloud points among bins in a way that
is not computationally complex. This is followed by Region-wise
Ground Plane Fitting, which is performed to estimate the partial
ground for each bin. Finally, Ground Likelihood Estimation is
introduced to dramatically reduce false positives. As experimen-
tally verified on SemanticKITTI and rough terrain datasets,
our proposed method yields promising performance compared
with the state-of-the-art methods, showing faster speed compared
with existing plane fitting–based methods. Code is available:
https://github.com/LimHyungTae/patchwork
Index Terms—Range Sensing; Mapping; Field Robots; Ground
Segmentation
I. INTRODUCTION
I
N recent years, there has been an increased demand to
perceive surroundings for mobile platforms, such as Un-
manned Ground Vehicles (UGVs), Unmanned Aerial Vehicles
(UAVs), or autonomous cars. To accomplish this, numerous
researchers have applied various 3D perception methods [1]–
[4]. In particular, a 3D light detection and ranging (LiDAR)
sensor has been extensively deployed due to allowing for
centimeter-level accuracy and omnidirectional sensing, as well
as its ability to measure great distances compared with stereo
cameras [1], [5], [6]. Accordingly, a 3D point cloud captured
by a LiDAR sensor is utilized for semantic segmentation [7],
[8], tracking [9], detection [10], and so forth.
In this paper, we specifically focus on ground segmentation
tasks [11], [12]. There are two main purposes of ground
segmentation. One is to estimate the movable area [3], [13] for
This work was supported by the Industry Core Technology Development
Project, 20005062, Development of Artificial Intelligence Robot Autonomous
Navigation Technology for Agile Movement in Crowded Space, funded by
the Ministry of Trade, Industry & Energy (MOTIE, Republic of Korea) and
by the research project “Development of A.I. based recognition, judgement
and control solution for autonomous vehicle corresponding to atypical driving
environment,” which is financed from the Ministry of Science and ICT
(Republic of Korea) Contract No. 2019-0-00399. The students are supported
by the BK21 FOUR from the Ministry of Education (Republic of Korea).
1
Hyungtae Lim,
1
Minho Oh, and
1
Hyun Myung are with the School of
Electrical Engineering, KI-AI, KI-R at KAIST (Korea Advanced Institute
of Science and Technology), Daejeon, 34141, South Korea. {shapelim,
minho.oh, hmyung}@kaist.ac.kr
GLE
Concentric Zone Model Output of GLE
Ground
Estimate
3D scan
For each bin
R-GPF
Fig. 1. Overview of our proposed method called Patchwork.
Patchwork mainly consists of three parts: Concentric Zone
Model (CZM)–based polar grid representation, Region-wise
Ground Plane Fitting (R-GPF), and Ground Likelihood Esti-
mation (GLE).
successful navigation. The other purpose, on which this paper
places more emphasis, is the segmentation of a point cloud
to recognize or track moving objects. Terrestrial vehicles or
humans inevitably come into contact with the ground [14];
ideally, dynamic objects can be recognized in a simple way,
such as through Euclidean clustering if the ground is well
estimated [8], [15]. Furthermore, because most cloud points
belong to the ground, ground segmentation can significantly
reduce computational power when one is performing object
segmentation or detection in a preprocessing stage [16]. Thus,
ground in this study refers to not only the road, which is a
movable area, but also all regions that moving objects can
come into contact with, including sidewalks or lawns.
In this study, as presented in Fig. 1, we propose a novel
Concentric Zone Model (CZM)–based region-wise ground
segmentation method, called Patchwork, which is an extension
of Region-wise Ground Plane Fitting (R-GPF) in our previous
study [14]. The aim of R-GPF in our previous study was to
estimate the ground points for static map building purposes,
whereas here, we focus only on ground segmentation on a
3D point cloud. We also conduct detailed experiments on the
impact of the bin size, which was not covered in our previous
paper.
In summary, the contribution of this paper is threefold:
• To the best of our knowledge, it is the first attempt to
analyze the impact of bin size when estimating ground
planes in complex urban environments using the Se-
manticKITTI dataset [1]. Accordingly, an efficient, non-
uniform, region-wise representation of a 3D point cloud
is proposed, referred to as a CZM–based representation
whose bin size is different depending on each zone.
• Also, we leverage Ground Likelihood Estimation (GLE)
in terms of uprightness, elevation, and flatness to deter-
arXiv:2108.05560v2 [cs.RO] 10 Mar 2022
2
mine whether each bin is ground.
• Our proposed method shows promising performance over
the state-of-the-art, region-wise fitting–based methods at
more than 40 Hz. In particular, Patchwork estimates
the ground points with the least recall variance, which
shows that our proposed method overcomes the under-
segmentation issue in complex urban environments.
II. RELATED WORKS
A. The Difficulties of Ground Segmentation
One may argue that it is a simple task that can be easily esti-
mated by filtering a point cloud based on sensor height or using
RANSAC [17] which is a renowned method for estimating a
plane. Unfortunately, there are three main issues that impede
algorithms from conducting precise ground segmentation: a)
there exists a partially steep slope or bumpy road, b) curbs
or flower beds make some regions uneven, and c) because all
surrounding objects are taken into account as outliers in the
ground segmentation tasks, these objects hinder plane fitting.
For these reasons, sometimes under-segmentation occurs, in
which case points belonging to different objects are merged
into the same segment [6], [12].
B. Ground Plane Estimation Methods
To tackle these issues, numerous researchers have studied
various approaches. For instance, Douillard et al. [4] and Chen
et al. [18] employed Gaussian process–based methods. On the
other hand, Tse et al. [19], Byun et al. [3], and Rummelhard
et al. [20] proposed Markov Random Field–based methods.
These methods can be used to estimate detailed ground points
yet requiring much computational time, so it may not be
appropriate to use them as preprocessing algorithms whose
speed should be guaranteed at more than 20 Hz.
C. Scan Representation
Meanwhile, grid representation–based methods have been
widely utilized to leverage expressibility compared with sin-
gular plane model–based methods [4], [9]. In particular, polar
grid representation, which treats a point cloud in cylindrical
coordinates, is commonly employed these days because it
naturally compensates for the geometric characteristics of
3D LiDAR sensors [11], [12], [14], [16], [21]. In practice,
Thrun et al. [5] presented a grid cell–based binary ground
classification method in a probabilistic way to predict the
movable area for autonomous driving in the DARPA challenge.
These methods are mainly divided into two categories: a)
elevation map–based and b) model fitting–based methods.
Accordingly, the latter category can be further classified into
two main methodologies: a) line fitting–based and b) plane
fitting–based methods.
D. Elevation Map–based 2.5D Grid Representation
First, elevation map–based methods are used to distinguish
between ground and non-ground points by encoding a 3D
point cloud into 2.5D grid representations [5], [9]. Thrun et al.
[5] utilized relative height and Asvadi et al. [9] used average
height and its covariance on each grid. These methods have
strong advantages over other methods in terms of speed and
computational cost. However, there are some potential risks
that sometimes a steep slope region could be considered as a
non-ground region because of large z value difference between
its supremum and infimum points with respect to Z-axis.
E. Multiple Line Fitting–based Ground Segmentation
Next, Himmelsbach et al. [11] and Steinhauser et al. [21]
introduced 2D line fitting on a uniform polar grid repre-
sentation to estimate the straight-line equation on each grid.
Then, in each grid, it was determined whether points were
ground points by comparing between constant thresholds and
the estimated parameters, such as the point-to-line distance,
gradient, or y-intercept.
F. Multiple Plane Fitting–based Ground Segmentation
Sharing their views of region-wise fitting yet improving ro-
bustness, other researchers have conducted region-wise plane
fitting–based approaches [8], [12], [14], [16]. For instance,
Zermas et al. [8] divided a point cloud into three parts
along the x-axis of the body frame, which is the forward
direction of a vehicle. This method is based on the premise
that a slope usually changes along the x-axis; however, this
assumption sometimes fails when it comes to a bumpy road
or a complex intersection. To resolve the problem, Narksri
et al. [12] proposed a slope-robust method using consecutive
ring patterns in the scan data as well as the concept of the
continuity of the region-wise estimated plane along the radial
direction. Furthermore, Narksri et al. [12] and Cheng et al.
[16] proposed an adaptive way of setting a grid size depending
on the density of the cloud points or the incidence angle.
G. Deep Learning-based Methods
Of course, as the deep learning era has come, Milioto et
al. [7] proposed RangeNet++ to estimate point-wise labels
on a 3D point cloud and Paigwar et al. [22] presented
GndNet, which estimates ground plane elevation information
in a grid–based representation to discern ground points in
real time. Unfortunately, these methods usually require high
computational resources. In addition, these methods tend to
be highly fitted to the environments of train dataset; thus, the
performance of those can be potentially degraded when used
in quite different environments from the training dataset or
different sensor configuration [23].
III. METHODOLOGY OF PATCHWORK
The following paragraphs highlight the problem definition
and the reasoning behind each module of Patchwork. Patch-
work mainly consists of three parts: CZM, R-GPF, and GLE.
A. Problem Definition
First, we begin by denoting a point cloud at the moment
as P. Then, let P = {p
1
, p
2
, . . . , p
k
, . . . , p
N
} be a set of
cloud points that contain N points at the moment acquired by
a 3D LiDAR sensor, where each point p
k
consists of p
k
=
{x
k
, y
k
, z
k
} in the Cartesian coordinates. In this paper, P is
definitely classified into two classes: a set of ground points,
G, and its complement, G
c
, which satisfy G ∪ G
c
= P. Note

剩余7页未读，继续阅读

身份认证 购VIP最低享 7 折!
30元优惠券
资源推荐
资源评论
良辰与日月
  • 粉丝: 135
  • 资源: 3
上传资源 快速赚钱
voice
center-task 前往需求广场，查看用户热搜

最新资源