移动机器人使用RGBD建图1
需积分: 0 183 浏览量
2022-08-03
21:01:12
上传
评论
收藏 4.58MB PDF 举报
Grid Map Guided Indoor 3D Reconstruction for Mobile Robots with
RGB-D Sensors
Boyu Zhang
1
, Xuebo Zhang
1
, Xiang Chen
2
, and Yongchun Fang
1
Abstract— This paper presents a novel automatic indoor
three-dimensional (3D) scene reconstruction approach which
is guided by a beforehand two-dimensional (2D) grid map.
The proposed system collects only a few images with a RGB-D
sensor (Kinect v2) at poses precalculated by the grid map to
reconstruct a complete model of the indoor environment. To
remove noises result from mirror reflection, a wall detection
based cloud filtering method is proposed as preprocessing for
the point clouds. Then, robot location information provided by
a laser scanner is utilized as an initial guess of the point cloud
registration which uses Iterative Closest Point (ICP). Finally, to
maintain global consistency and improve the overall accuracy, a
graph optimization strategy is applied. Experimental results are
provided to show the effectiveness of the proposed approach.
Key words–3D reconstruction, RGB-D sensor, mobile robots,
graph optimization
I. INTRODUCTION
Three-dimensional (3D) construction is widely used in
many applications such as virtual reality, medical imag-
ing, industrial manufacturing and so on. It is also an im-
portant part in 3D simultaneous localization and mapping
(SLAM) [1], [2], [3], since an accurate 3D model of the
indoor environment can provide much useful information.
In the past, the acquisition of 3D models relies on sensors
like 3D LIDAR scanners which are quite expensive, or
2D cameras which is lack of scale information. In recent
years, the appearance of the cheap Microsoft Kinect sensor
contributes a lot to the 3D reconstruction tasks [4]. The new
generation of Kinect device, Kinect v2, is even much better
than Kinect v1 in both resolution and accuracy. More and
more attention is paid to 3D reconstruction [5], [6], [7], [8].
Researchers in this filed can be divided into two groups,
the one working on computer vision and the one studying
SLAM.
Researchers working on computer vision focus on restor-
ing the real scene accurately in details with a passive method
and reconstruct a vivid model. There are many works with
impressive performance [9], [10], [11], wherein the most
representative work is the KinectFusion [12], [13] allowing
real-time dense volumetric reconstruction. This algorithm
fuses all the depth data streamed from a Kinect sensor
together and create a model of high precision. However,
∗ This work is supported in part by National Natural Science Foundation
of China under Grant 61573195 and U1613210.
1
Boyu Zhang, Xuebo Zhang (corresponding author) and Yongchun Fang
are with the Institute of Robotics and Automatic Information System
(IRAIS), and also Tianjin Key Laboratory of Intelligent Robotics (TJKLIR),
Nankai University, Tianjin, China, 300071
2
Xiang Chen is with the Department of Electrical and Computer Engi-
neering, University of Windsor, Ontario, Canada, N9B3P4.
as only Iterative Closest Point (ICP) algorithm is used as
the odometry, the system is not robust enough and easily
fails when camera displacement is large between frames.
Also, the implementation is mainly aimed at reconstructing
some objects or part of indoor areas and may be unsuitable
for large, complete indoor environments due to the drift of
camera poses. With the growth of the model size, the problem
of inconsistency will become more and more obvious.
Researchers in RGB-D SLAM focus on automatic re-
construction of large-scale indoor environments [3], [14],
[15]. The demand for map equality in SLAM is not as
high as KinectFusion, but the global consistency of the
map is considered. Usually, they conduct a loop closure
detection and keep the consistency of the map with a graph
optimization strategy [16], [17]. When the robot reaches a
place that has been visited, a closed loop will be added to
the pose graph of the robot. This approach can solve the
problem of camera pose drift as it updates the pose graph
from the global point of view. As for camera localization, in
RGB-D SLAM, a combination of feature matching and ICP
is often adopted as the odometry. ORB [18] or SIFT [19]
features of the images are extracted and matched to provide
a coarse point cloud registration, and then the result is refined
with ICP. Remarkable performance can be achieved, but in
fact there are still problems to be addressed. On the one
hand, feature extraction of high quality and accuracy is a
challenging problem under changing illumination, scale and
perspective transformation. Inaccurate features may have a
negative effect on the overall mapping performance. On the
other hand, once the robot arrives at some places lack of
features, it easily gets lost and the mapping process will fail.
Therefore, more reliable methods are still needed.
Another problem involved in this article is noise treatment.
As an important part of 3D reconstruction, point cloud
filtering is also a concern in a lot of works [20], [21]. In this
paper, the Kinect v2 sensor is utilized for its high resolution
and good accuracy. However, using the method of Time of
Flight (TOF), the Kinect2 sensor provides erroneous mea-
surement when there is mirror reflection in the environment.
Traditional filtering methods such as bilateral filtering and
statistical outlier removal are aimed at smoothing the data
and removing static outlier, and these methods are unsuit-
able for the clustered noises result from mirror reflection.
Therefore, a new filtering method is needed to handle this
problem, in indoor environments.
In this paper, we propose a method of building a 3D map
guided by a 2D grid map which is obtained in advance
by running the Karto-SLAM algorithm [22]. Based on the
Proceedings of the 2018 IEEE/ASME International
Conference on Advanced Intelligent Mechatronics (AIM),
Auckland, New Zealand, July 9-12, 2018
WAT3.1
978-1-5386-1854-7/18/$31.00 ©2018 IEEE 498
在2D图上重建3D，使用kinect v2
评论0