IEEE Robotics & Automation Magazine - March 2023 - 43

future state of these semistatic map
points and obtain the predicted map that
is more suitable to the current environment.
In addition, our method is also better
than other map prediction-based
methods because our method considers
the occlusion and error detection of the
map points, and the GMP is more accurate.
The absolute trajectory error (ATE)
of our system is also less, which means
better localization performance in the
simulation scene.
"
OUR METHOD IS MORE
ACCURATE COMPARED TO
THE METHODS THAT USE
THE STATIC-WORLD
ASSUMPTION.
„
TUM DATASET EVALUATIONS
We use TUM datasets to demonstrate the effectiveness of
our system on dynamics removal. TUM datasets [18] are the
indoor sequences from the TUM RGB-D benchmark. The
TUM datasets contain both the static and dynamic scenarios
in indoor environments. We divided the environments of
the TUM benchmark into three types: static, low dynamic,
and high dynamic according to the part of the FoV covered
by moving objects. The low-dynamic environment sequences
are the sitting and desk-person sequences, and the highdynamic
environment sequences are the walking sequences.
We run our dynamic MPR method on sequences of the
TUM datasets and measure the ATE of different methods.
Our method is performed on a desktop
computer equipped with an Intel Core
i7-8750 (2.2-GHz) CPU and 8 GB of
random-access memory. For comparison,
our results contain the following
state-of-the-art methods: ORB-SLAM2
[14], ORB-SLAM3 [15], dense VO CPU
[17], Depth-Edge [1], DynaSLAM [2],
and Point-Correlations [5]. Results of
the ATE comparison of our method and
other visual SLAM methods are shown
in Tables 2 and 3, respectively.
For low-dynamic environments, most of the influence of
moving objects can be eliminated. Our method is more accurate
compared to the methods that use the static-world assumption
[14], [15]. In addition, our system is also better than
the methods that consider moving objects. Those methods
mistakenly remove some static objects or aggressively segment
images. Learning-based methods [1], [2] mask the entire
image region of the person. However, most of the body is static
in low-dynamic environments. As a result, less environment
information is used by these methods, and the number of
available static map points for pose estimation is not enough,
resulting in poor localization performance.
For high-dynamic environments, the features on people's
bodies move rapidly and should be removed. Our method is
RoboSense
RS-LiDAR-16 Lidar
Intel RealSense
D435i Camera
(a)
(b)
13
13
5
7
24
(c)
24
(d)
FIGURE 6. The experiment scenarios for our intelligent wheelchair. (a) The simulation of our JiaoLong intelligent wheelchair. (b) The
prototype of our JiaoLong intelligent wheelchair. (c) Sample images of our simulation scenes. (d) Images of the real-world experiment
scene under different time and lighting conditions. The upper and lower images in each column show approximately the same place in
different data sequences, i.e., the scene changes. (c) Images of gazebo datasets and (d) images of real-world datasets.
MARCH 2023 IEEE ROBOTICS & AUTOMATION MAGAZINE
43
6
8

IEEE Robotics & Automation Magazine - March 2023

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