IEEE Robotics & Automation Magazine - March 2023 - 37

reality. Therefore, there are increasing demands for accurate
and robust visual SLAM systems to operate in real-world
environments over extended periods of time. However, most
of the existing visual SLAM systems assume a static environment
and build a constant map, which is unsuitable for
dynamics in environments. Long-term
visual SLAM (LTVS) is becoming
intractably difficult due to the many
challenges in dynamic environments and
long-term operations.
Many existing visual SLAM systems
are usually designed and evaluated for a
single operation. However, in real-world
deployment, robots usually operate at
a region day after day, with the requirement
of reusing an accurate map in each
operation. As the scene changes in real
life, the robots have to face the fact that
the previous observations are incomplete
and uncertain, leading to poor and incorrect
performance in localization.
To eliminate the influence of dynamic objects, we classify
"
a persistent filter to recognize dynamic map points in environments
and remove them. Furthermore, some methods estimate
both the motion of a camera and dynamic objects. Huang et
al. [7] offer ClusterVO, a stereo visual odometry that simultaneously
optimizes the poses of the camera and multiple
moving objects, regarded as clusters of
point landmarks.
For semistatic objects in real-world
THE MAP POINTS
MATCHING MODULE IS
USED TO OBTAIN AN
ACCURATE TIME SERIES
OF MAP POINTS.
„
most dynamics into two types:
1) Dynamic: These types of objects rarely appear and move fast,
such as humans moving around. These dynamics usually
occur in a short timescale during a single robot traversal.
2) Semistatic: These types objects move slowly or seldom
change their position. These dynamics usually happen
across a longer timescale that can only
be detected by multiple visits to the same
place. In this article, we focus on both
dynamic and semistatic objects in long-term
environments and eliminate their influence
using different methods.
For dynamic objects in real-world environments,
removing the dynamic parts is one of the
possible approaches to achieving reliable localization.
Li and Lee [1] present the Depth-Edge
SLAM system, which uses only depth edge points
for frame-to-KeyFrame registration. Bescos et al.
[2] propose DynaSLAM, a visual SLAM system that
adds the capabilities of dynamic object detection and
background inpainting. Bahraini et al. [3] offer a novel
approach to segment and track multiple moving objects
in real dynamic environments, and detected objects are
classified into stationary and moving objects using the
Multilevel-Random Sample Consensus algorithm. Kim
and Kim [4] present a novel static point-cloud map construction
algorithm called Removert for use within dynamic
urban environments. Leaving only static points and excluding
dynamic objects is a critical problem in various robust robot
missions in changing outdoor environments. Dai et al. [5] propose
a point-correlation SLAM system that utilizes the correlation
between map points to separate static and dynamic
points and create a sparse graph using Delaunay triangulation.
Hashemifar and Dantu [6] present the practical application of
environments, there are usually two types
within a scene: periodic and abrupt.
Periodic changes can be both natural,
e.g., sunlight changes, or artificial, e.g.,
turning on or off lights regularly. Periodic
changes include changes that occur
slowly and continuously, e.g., the weakening
of lighting conditions. Abrupt changes
usually occur once and drastically
change the environment. The solution for
these semistatic objects is to predict the
changes in the environment. In [8] is our
prior work, where we make predictions
for some semistatic objects with a fixed set of states, such as
the " opened " and " closed " states of the door. Then we use the
predicted map fused with current observations to maintain a
global map in the changing environment. In [9] and [10], a new
approach is presented for long-term mapping in a dynamic
indoor environment. The authors use a spectral model to represent
arbitrary timescales of environments' dynamics with low
memory requirements. Their system can accurately predict the
future state of the map point when facing periodic changes. In
[11] is a recent work that models persistence in semistatic environments
and incorporates the feature persistence estimation
into graphical mapping techniques.
In addition, topological map-based methods are another
approach for LTVS in dynamic environments. In [12], the
authors propose a topological, local-metric framework, aiming
to deal with environmental changes and erroneous measurements
to achieve constant complexity. They organize the sensor
data collected by the robot in a topological graph, and the geometry
is encoded only in the edge to relax global consistency.
In this article, we propose an LTVS system with dynamic
map point removal and global map prediction (GMP) (see
Figure 1). Our system is based on an ORB-SLAM [14], [15]
system with the BPF [11]. For the dynamic objects in the environment,
we use the persistence filter to remove dynamic map
points. As for the semistatic objects in the environment, we
propose a GMP method to model the time series of these map
points and predict their future states. To obtain a more accurate
time series of map points, a map point matching algorithm
is proposed, which fuses 2D pixel-metric information and 3D
voxel information, making it more efficient and accurate.
Then, by using the BPF and survival analysis, each map point
in the temporal map is classified into static and semistatic
based on their persistence probability. The future state of each
semistatic map point is predicted according to its previous
observations, and the predicted maps are built for localization.
MARCH 2023 IEEE ROBOTICS & AUTOMATION MAGAZINE
37
IMAGE LICENSED BY INGRAM PUBLISHING
IEEE Robotics & Automation Magazine - March 2023

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