IEEE Robotics & Automation Magazine - March 2023 - 44

better than the methods that assume a static world. When
compared with those learning-based methods, our method is
not good enough. Those methods use a human-shape prior,
and the entire image region of the person is masked using
learning techniques; they are specifically designed for a
dynamic TUM dataset. Those learning-based methods utilizing
only static information perform better in high-dynamic
sequences. In a walking-static sequence, our method perTABLE
1. A comparison of the absolute trajectory error
in different datasets.
METHODS
ORB-SLAM2
ORB-SLAM3
DVO SLAM
Depth-Edge
DynaSLAM
Point-Correlations
RMSE
MAX
forms better than other methods. In this sequence, the camera
position is relatively static, and our method can obtain
more accurate poses by removing the dynamic map points
and KeyFrames.
DATASETS: LOW DATASETS: MEDIUM DATASETS: GAZEBO
RMSE MAX
RMSE
0.4294 1.0744 0.5827
0.4162 0.6578 0.4728
0.7342 1.6831 0.8754
0.4327 0.9792 0.5342
0.4037 0.9062 0.5061
0.3931 0.6972 0.4521
LTVS-BPF without MPR 0.4221 0.9197 0.4724
LTVS-BPF without GMP 0.4323 0.9124 0.4533
LTVS-BPF without OD 0.4778 1.4021 0.5659
LTVS-BPF without 3DM 0.4342 1.0231 0.4625
LTVS-BPF
MAX
1.3235 0.1242
0.8865 0.0684
1.8754 0.1562
1.5839 0.1048
1.2672 -
0.4061
0.1392
0.685
0.3741 0.7063 0.4324
RMSE: root-mean-square error; Max: maximum; DVO: dense VO.
0.8963 0.0504
0.9721 0.0624
0.8876 0.0682
1.7435 0.1321
0.9531 0.0663
0.8021 0.0551
REAL-WORLD EXPERIMENTS
In this section, we introduce our real-world experiments platform
and datasets. The JiaoLong intelligent wheelchair is
used as our experiments platform. It is
equipped with RoboSense RSLiDAR-16
lidar and an Intel RealSense
D435i camera. An industrial PC MIC770
with an Intel Core i7-9700TE CPU
is also employed, with an Ubuntu and
Robotic Operating System. To demonstrate
the effectiveness and general
applicability of our system, we collect
our own real-world datasets. The
experiment scenarios and the detailed
information of our intelligent wheelchair
are shown in Figure 6.
For the real-world datasets, we col0.3245
-
0.1467
0.1895
0.1963
0.4156
0.1593
0.1355
lected
datasets from November 2021 to
February 2022, from 9 a.m. to 9 p.m.
(where the interval is 3 h), seven days
per week. Our wheelchair was operated
day after day in an indoor hall
environment. During data collection,
there were different types of dynamic
TABLE 2. A comparison of the ATE on the TUM benchmark in low-dynamic environments.
THE RMSE OF TRAJECTORY ALIGNMENT (M)
SEQUENCES
Low
dynamic
fr2/desk_with_person 0.0062
fr3/sitting_static
fr3/sitting_xyz
0.0073
0.0111
fr3/sitting_halfsphere 0.0234
fr3/sitting_rpy
0.0211
RMSE: root-mean-square error; DVO: dense VO.
ORB-SLAM2 ORB-SLAM3 DVO SLAM DEPTH-EDGE DynaSLAM
0.0068
0.1037
0.0072
0.0096
0.0247
0.0203
0.0484
0.0119 -
0.242
0.2198
0.1756 -
0.0397
0.0432
-
-
0.015
0.017
-
POINTCORRELATIONS
LTVS-BPF
0.0075
0.0091
0.0096
0.0235
0.0225
0.0063
0.0061
0.0087
0.0168
0.0127
TABLE 3. A comparison of the ATE on the TUM benchmark in high-dynamic environments.
THE RMSE OF TRAJECTORY ALIGNMENT (M)
SEQUENCES
High dynamic fr3/walking_halfsphere 0.3666
fr3/walking_static
fr3/walking_xyz
fr3/walking_rpy
0.0331
0.4216
0.1861
RMSE: root-mean-square error; DVO: dense VO.
ORB-SLAM2 ORB-SLAM3 DVO SLAM DEPTH-EDGE DynaSLAM
0.025
0.3165
0.0283
0.3157
0.1678
1.0136
0.7515
1.383
1.2922
0.0489
0.0261
0.0601
0.1791
0.0108
0.015
0.035
POINTCORRELATIONS
LTVS-BPF
0.0354
0.0128
0.0874
0.1608
0.171
0.0106
0.1583
0.1121
44 IEEE ROBOTICS & AUTOMATION MAGAZINE MARCH 2023
IEEE Robotics & Automation Magazine - March 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - March 2023
