Article

LIO-SAM++: A Lidar-Inertial Semantic SLAM with Association

Optimization and Keyframe Selection
Bingke Shen 1,2 , Wenming Xie 1,2, *, Xiaodong Peng 1,2,3 , Xiaoning Qiao 1,2 and Zhiyuan Guo 1,2

1 National Space Science Center, Chinese Academy of Sciences, Beijing 100190, China;
[email protected] (B.S.); [email protected] (X.P.); [email protected] (X.Q.);
[email protected] (Z.G.)
2 University of Chinese Academy of Sciences, Beijing 100049, China
3 Hangzhou Insititute for Adcanced Study, Chinese Academy of Sciences, Hangzhou 310024, China
* Correspondence: [email protected]

Abstract: Current lidar-inertial SLAM algorithms mainly rely on the geometric features of the lidar
for point cloud alignment. The issue of incorrect feature association arises because the matching
process is susceptible to influences such as dynamic objects, occlusion, and environmental changes.
To address this issue, we present a lidar-inertial SLAM system based on the LIO-SAM framework,
combining semantic and geometric constraints for association optimization and keyframe selection.
Specifically, we mitigate the impact of erroneous matching points on pose estimation by comparing
the consistency of normal vectors in the surrounding region. Additionally, we incorporate semantic
information to establish semantic constraints, further enhancing matching accuracy. Furthermore, we
propose an adaptive selection strategy based on semantic differences between frames to improve the
reliability of keyframe generation. Experimental results on the KITTI dataset indicate that, compared
to other systems, the accuracy of the pose estimation has significantly improved.

Keywords: lidar-inertial SLAM; semantic information; association optimization; keyframe selection

1. Introduction
Citation: Shen, B.; Xie, W.; Peng, X.;
Qiao, X.; Guo, Z. LIO-SAM++: A SLAM (Simultaneous localization and mapping) technology involves a robot con-
Lidar-Inertial Semantic SLAM with structing environment maps and estimating its position in an unknown environment
Association Optimization and by collecting data from its surrounding environment using its mounted sensors. Due
Keyframe Selection. Sensors 2024, 24, to the continuous development of intelligent technology, SLAM technology has gained
7546. https://doi.org/10.3390/ widespread attention and application in robotics and autonomous driving. SLAM can
s24237546 be categorized into Lidar-based and visual-based approaches, depending on the type of
Academic Editor: Hui Kong
sensors used for environmental perception. Visual SLAM captures abundant visual data
through cameras, providing color and texture information of the scene; however, it is
Received: 24 September 2024 susceptible to changes in illumination. Conversely, LIDAR (Light Detection and Ranging)
Revised: 8 November 2024 is capable of accurately measuring distance and remains unaffected by varying illumination
Accepted: 25 November 2024 conditions, rendering it more commonly utilized in large-scale environments. Notably,
Published: 26 November 2024
the laser-based multi-sensor fusion can further enhance environmental perception and
serve as a crucial method for improving the performance of SLAM systems.
In recent years, many solutions have been proposed for lidar-inertial SLAM, such as
Copyright: © 2024 by the authors.
LEGO-LOAM (Lightweight and Ground-Optimized Lidar Odometry and Mapping) [1],
Licensee MDPI, Basel, Switzerland. LIO-SAM (Lidar Inertial Odometry via Smoothing and Mapping) [2], and FAST-LIO (A
This article is an open access article fast robust LiDAR-inertial odometry) [3]. These methods perform well in some scenes
distributed under the terms and but may be disturbed in the real word. Most of these works select the associated points
conditions of the Creative Commons in the feature-matching process based only on the nearest-neighbor distance information.
Attribution (CC BY) license (https:// However, this way of association does not consider the effects of point positions and
creativecommons.org/licenses/by/ densities between point cloud frames, which results in a lack of robustness in the pose
4.0/). transformation. In addition, the point cloud maps of the environment generated by these

Sensors 2024, 24, 7546. https://doi.org/10.3390/s24237546 https://www.mdpi.com/journal/sensors

Sensors 2024, 24, 7546 2 of 18

algorithms only contain geometry information and lack semantic descriptions of the objects
in the scene, making it difficult to understand the scene in depth. As neural networks
advance in the field of point cloud semantic segmentation, researchers have integrated
semantic information into LIDAR-based SLAM frameworks. Semantic-aided LiDAR SLAM,
such as SUMA++ (Efficient LiDAR-based Semantic SLAM) [4] and SA-LOAM (Semantic-
aided LiDAR SLAM) [5], use semantic segmentation networks to obtain the semantic labels
of the scene, which provides additional constraints for the point cloud alignment process
and reduces the dependence on the environment geometry. Although these methods
have made some progress, the performance of these can still be deficient when facing
environmental changes in the scene.
In this paper, we propose a lidar-inertial SLAM with association optimization and
keyframe selection. We optimize the feature association by combining geometric and seman-
tic information and reduce trajectory drift by adaptively generating keyframes according
to scene changes. The main contributions presented in this work can be summarized
as follows:
(1) We propose a feature association method based on neighborhood normal vector con-
sistency, which uses local region geometric information to filter the nearest neighbor
feature points to reduce the problem of inaccurate feature point association in point
cloud registration.
(2) We propose a semantic-assisted point cloud registration method for pose estimation,
which constructs a weighted cost function based on semantic attributes to achieve
more reliable odometry data association and improve pose estimation accuracy.
(3) We propose a semantic-based keyframe optimization method to adaptively generate
keyframes through the difference of semantic information between frames, which
reduces the probability of losing the valid point cloud frame.

2. Related Work
2.1. Feature Association
In SLAM systems, the process of scan registration is obtaining the pose transformation
by aligning two successive point cloud frames or the current point cloud frame with the
established map. Geometry-based association methods often employ geometry primitives,
such as points, lines, planes, or curvature features, to facilitate rapid matching. In LOAM [6],
point cloud curvature is computed to categorize points as either surface or corner features.
Residual equations are then formulated from the distances of corner points to lines and
surface points to planes, which are utilized to optimize pose estimation via the Levenberg-
Marquardt method. Many subsequent algorithms continue to adopt this feature-matching
approach, such as LeGO-LOAM [1], which reduces the complexity of computation by
segmenting the ground points of the point cloud frames and then clustering them to
obtain labels, and using the label information as the constraints for inter-frame matching.
LIO-SAM [2] extended LeGO-LOAM by removing the frame-to-frame matching part and
incorporating IMU pre-integration and GPS measurements factor within the factor graph
optimization. FAST-LIO [3] also extracts the corner and surface point features and proposes
a new Kalman gain formula to achieve efficient computation. In an effort to further
improve the reliability of solving pose transformation, some works began to focus on
the categorization of features. Guo et al. [7] introduced an approach using Principal
Components Analysis (PCA) to distinguish between corner and surface points utilizing
the properties of the points. Similarly, MULLS [8] extracted various features based on
PCA, which proposed a multi-metric linear least squares method for iterative optimization
of pose estimation. Meng et al. [9] proposed an adaptive feature extraction method for
different scenes, which adjusts planar and linear feature thresholds according to local
environmental properties and constructs multiclass cost functions for distinct categories
of features.
The above research enhanced the precision of feature point association by catego-
rizing feature points and seeking within the identical category. However, we consider
Sensors 2024, 24, 7546 3 of 18

that feature points of the same class are still associated with nearest neighbors, which
affects the accuracy of the association in cases when point cloud frames and maps do not
overlap completely. Therefore, unlike these feature association optimization methods, our
approach improves the reliability of feature association by expanding the scope of the
nearest neighbor search and combining it with normal vector consistency for filtering.

2.2. Pose Estimation Combining Semantic Constraints

Semantic information plays a crucial role in environment understanding. High dimen-
sional semantic information can provide more a priori conditions for the system, which
can construct constraints that help optimize the state estimation and achieve more accurate
point cloud frame-to-map matching. SuMa++ [4], based on the SuMa, introduces semantic
information to optimize the pose estimation by adjusting the residual weights using se-
mantic consistency. When the semantic labels at the corresponding positions in the frame
and the map are the same, the residual weight is increased to improve the registration
robustness. Similarly, Zhao et al. [10] and Du et al. [11] proposed further use of geometry
features to construct residuals only if the semantic labels are identical.
These studies assign weight coefficients to the associated point pairs based on semantic
label consistency, thereby enhancing the precision and stability of SLAM and showing the
significant potential of using semantic information as a priori conditions. In contrast to
the methods, our approach considers the challenge of inaccurate semantic segmentation
in boundary areas and avoids complete dependence on the semantic labels of associated
feature points.

2.3. Keyframe Selection

SLAM systems are often required to process a vast quantity of lidar data, resulting in
significant algorithmic computational load due to the rapid lidar point cloud updates and
the information overlap between frames. To solve the problem, only keyframes instead of
all frames are usually utilized in the back-end optimization process, thus ensuring accuracy
while improving computational efficiency. Therefore, a reasonable keyframe selection
criterion is necessary. Currently, the criteria used to select keyframes may be classified into
two principal categories: heuristic-based methods and information theoretic methods.
Heuristic-based methods usually determine keyframes by time or pose changes. For ex-
ample, LIO-SAM [2] and SVO [12] calculate the translation distance and rotation Angle
of the robot pose estimation between the current frame and the last keyframe, selecting
the current frame as a keyframe upon exceeding a defined threshold of change. ORB-
SLAM [13] extracts the keyframes by checking the number of frames between two frames.
DSO [14] extracts the new keyframes by combining the field of view, camera position,
and exposure time variations as criteria for selecting new keyframes. VINS-MOON [15]
employs another heuristic strategy for keyframe selection, which is detecting whether the
feature count in the current frame falls within a set threshold range. In contrast, information
theoretic methods have more stringent metrics. These methods decide whether to insert a
new keyframe by evaluating the uncertainty or quality of the pose estimate of the current
frame. Kuo et al. [16] select keyframes based on the differential entropy reflecting the
uncertainty of the current camera estimation of the pose concerning the historical keyframe.
Jiao et al. [17] select keyframes by computing the information matrix and assessing the con-
vergence of the state estimation problem. Based on this work, InfoLa-SLAM [18] considers
the information between neighboring frames and selects keyframes more comprehensively
by constructing the Fisher information matrix.
Although the above keyframe selection methods are effective in reducing redundancy,
they primarily depend on geometric information or the outcomes of geometric optimization.
This approach will overlook the characteristics of scene changes. As a result, it may fail
to capture significant details, potentially introducing bias into the estimated trajectories.
Considering semantics can describe the scene effectively, we use the semantic information
to focus on the scene changes, generating more reasonable keyframes.
Sensors 2024, 24, 7546 4 of 18

3. Method
The architecture of the proposed lidar-inertial semantic SLAM LIO-SAM++ is depicted
in Figure 1. The input is point cloud data from the LiDAR and high-frequency pose
data from the IMU, and the output is the state estimation and an environment map with
semantic information. This process has several stages, including IMU pre-integration, lidar
odometry, keyframe selection, loop closure detection and graph optimization.
(1) During the IMU pre-integration phase, the IMU measurements between two frames
of LiDAR data are pre-integrated;
(2) During the lidar odometry phase, semantic segmentation is applied to the point cloud
frames via a pretrained network to get semantic labels; Subsequently, corner points and
surface points are extracted from these frames, with the IMU pre-integration providing
an initial pose estimate. The pose of the current frame is obtained by minimizing the
residuals with semantics and geometry. For details, refer to Section 3.1;
(3) During the keyframe selection phase, keyframes are generated based on the scene
semantic changes and estimated pose transformation for back-end graph optimization.
For details, refer to Section 3.2;
(4) During the loop closure detection phase, potential loop closure relationships are found
by detecting historical keyframes to establish loop closure constraints;
(5) During the graph optimization phase, lidar odometry factors, IMU pre-integration
factors and loop closure detection factors are constructed. These factors are used
to update the state of system estimation through factor graph optimization and the
output of accurate pose estimation at the rate of the IMU. Meanwhile, global consistent
point cloud maps are generated by transforming the keyframe collection to the world
coordinate system and fusing them with the constructed global map;
In the next sections, we present detailed information about each module.

Figure 1. The structure of the LIO-SAM++.

3.1. Lidar Odometry

The lidar odometry achieves the registration of a point cloud by fusing the lidar and
IMU data and then estimates the relative pose. Given a LiDAR point cloud pt at time
stamps t and the local map point cloud Mt the transformation matrix T between the point
cloud frame and the local map is estimated by feature matching.

T = arg min ∑ ∥mtj − T pit ∥2 (1)

T i

where mtj and pit is the corresponding points between the local map and point cloud frame.

3.1.1. Feature Extraction

We incorporate a semantic segmentation network for point clouds into the SLAM
framework to obtain semantic information. The goal of semantic segmentation of point
clouds is to categorize individual points, thereby getting their semantic labels. Considering
adapting to the frame rate requirements of sensors and meeting real-time requirements,
in this paper, we use RangeNet++ [19] to segment the de-distorted point cloud frames
Sensors 2024, 24, 7546 5 of 18

to obtain semantic information in the environment. RangeNet++ was trained on the

SemanticKITTI dataset [20] and is capable of distinguishing between 22 different semantic
categories, including roads, vegetation and buildings, among other objects.
We extract features from the segmented current lidar point cloud frames to describe
their geometric attributes, distinguishing corner and surface feature points based on cur-
vature calculations [2]. To avoid the clustering of feature points, the lidar scan line is
segmented into several parts, ensuring that points within proximity on the same scan line
are not processed repeatedly. The curvature c is obtained by calculating the spatial variation
of the surrounding points:

1
| S | · ∥ri ∥ j∈∑
c= ∥ (ri − r j )∥2 (2)
S,j̸=i

where S represents the collection of points located on either side of the current point along
the same lidar scan line, ri and r j denote the vectors from the origin of the coordinate
system to the points pi and p j . A point is designated as a corner feature if its curvature
exceeds a predefined threshold, whereas a point below this threshold is classified as a
surface feature.
Ultimately, the extracted corner features Fie and surface features Fis from the point
cloud of each frame, which are combined into a feature set Fi = { Fie , Fis }. This set represents
the current frame and is utilized for subsequent registration of point clouds.

3.1.2. Feature Association

The accuracy of pose estimation of SLAM depends on the accuracy of point cloud
matching, so correct feature association is crucial. Traditional feature association methods
usually directly obtain the nearest neighbors of corner and surface feature points and
use these points to construct linear and planar constraints for nonlinear optimization.
However, this assumption that the nearest neighbors are the associated feature points is not
completely reliable because it does not guarantee that the nearest neighbors are the correctly
associated points. The reason for this problem is that the search for nearest neighbors as
associated points requires that the corresponding points nearly overlap with each other.
In contrast, the transformed point cloud and the local maps cannot satisfy the nearly
overlapping condition due to the initial value impact and the difference in point density.
Erroneous point associations can adversely impact the precision of matching results, while
conventional methods often overlook the geometry structural information in point clouds.
However, this geometric information is crucial to find the correct association between
points. Therefore, we propose to use geometric information for feature point selection
by modeling the geometric distribution around each feature point and introducing a
feature association approach based on local normal vector description to find more accurate
associated features.
Firstly, the feature point set Fi from the current point cloud scan is mapped into
the world coordinate system using the pose transformation Tiimu obtained from IMU pre-
integration. In the transformed coordinate system, for each feature point f i , 15 candidate
matches MCi are found through nearest neighbor search in the feature point set Mi of the
local map. Next, the set of nearest points around the feature point f i is searched in the
point cloud scan to form a neighborhood R to quantify the geometric similarity between
the points. Finally, obtain the set of 5 nearest neighbor points MDi by filtering based on the
geometric similarity and used to construct the residual optimization to ensure the point
cloud matching accuracy better.
To quantify the geometric similarity between points, we propose a method for describ-
ing local areas that utilize a normal vector. The method calculates normal vectors through
the covariance matrix and thus combines the nearest distance information to achieve the as-
sociation of feature points. For each feature point f i in a point cloud, we first determine the
average position f¯ of the points within the neighborhood R. Subsequently, leveraging the
Sensors 2024, 24, 7546 6 of 18

positional variance of these neighboring points from the mean, we calculate the covariance
matrix C. Finally, we perform an eigendecomposition of C to extract its eigenvalues λ and
corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue is
the normal vector of the local region. The unit normal vector n̂ is defined by the formula:
N
1
f¯ =
N ∑ fi (3)
i =1
N
1
C=
N ∑ ( fi − f¯)( fi − f¯)T (4)
i =1

Cn = λmin n (5)
n
n̂ = (6)
∥n∥
where f i represents the coordinates of the ith point in the neighborhood, N is the total
count of points in the neighborhood, and λmin refers to the eigenvalue of the smallest
eigenvalue,n denotes the normal vector.
According to Equation (6), we compute the local unit normal vector n̂1 for the points
in the point cloud scan and the local unit normal vector n̂2 for each candidate matching
point in the map. Subsequently, we compare the difference between n̂1 and n̂2 . If the local
normal vector direction of a point in a point cloud scan is consistent with that of a matching
point in the local map, it indicates that the two points may lie on the same plane, and the
match is retained. The consistency of normal vectors is defined by the angular difference
between the vectors:
θ = arccos(n̂1 · n̂2 ) (7)
where n̂1 and n̂2 are the unit normal vectors corresponding to the point cloud scan and the
local map, respectively. Consistency of the normal vector is indicated if the angle θ is below
a predetermined threshold.By calculating the local unit normal vector consistency between
the point cloud frames and the corresponding points of the local map, five more accurate
matching points are selected from the 15 candidate matching points. These five matching
points will be used to construct geometric residual and semantic attribute constraints to
improve the reliability of the matching results.
The feature association strategy based on normal vector consistency has higher robust-
ness than the traditional method that only relies on nearest neighbor search. The traditional
methods may produce false matches due to noise. In contrast, the proposed method can
effectively filter out incorrect matches and preserve points with similar geometric features,
thus improving the accuracy of the matches.
The process of the association point selecting method is shown in the Algorithm 1.

3.1.3. Pose Estimation

In SLAM algorithms, pose estimation is performed by optimizing the constraint resid-
uals of the associated feature points, and the calculation results are directly related to the
accuracy of the final pose estimation. Traditional constraint construction methods often
rely on the established feature association relationships on which constraint residuals are
computed separately for the associated corner and surface feature points to minimize the
geometric error between the point cloud and the map. However, the association relation-
ships of feature points in these methods are susceptible to occlusion or dynamic objects,
which reduces the accuracy of the associations. To reduce the impact of mismatching
on the optimization process, this paper introduces semantic information to assist map
matching during the geometric constraint construction process, thereby improving the
robustness of the optimization process. The method uses semantic information to provide
an additional basis for the judgment of matching pairs, which can reduce the impact of
semantic inconsistency points on the optimization.
Sensors 2024, 24, 7546 7 of 18

Algorithm 1 Feature association based on neighborhood normal vector consistency

input: current point cloud frame set L = { p1l , p2l , ..., plm }, local map point cloud set
M = { p1m , p2m , ..., pm
n }, the pose transformation T obtained from IMU pre-integration.
output: The set Ps of 5 points that satisfy the requirement
1. for pil in L do
2. lidar2map(pil ,pil →m )
3. M15 ← kdMapNearest(pil →m ,15)
4. for pim in M15 do
5. M10 ← kdMapNearest(pim ,10)
6. L10 ← kdLidarNearest(pil →m ,10)
7. if(checkNormalConl(L10 ,M10 ,T))
8. add pim to Ps
9. if(Ps .size==5)
10. break;
11. end if
12. end if
13. end for
14. end for

The method flow is as follows: firstly, the distance residuals are constructed separately
for corner and surface feature points. Next, the semantic consistency probability is calcu-
lated and the residuals are weighted using this probability. Finally, the weighted residuals
are optimized as the cost function of the lidar odometry.
For corner points, the residual is defined as the distance from the feature point of the
current frame to the line formed by the 2 nearest neighbors of the corresponding 5 points in
the local map. For surface points, the residual is defined as calculating the distance from the
feature point of the current frame to the plane formed by the 3 non-collinear points among
the corresponding 5 points in the local map. The corner point residuals are calculated
as follows:
e − f e ) × ( f e − f e )|
|( f i,k u,m i,k v,m
de = e − fe |
(8)
| f u,m v,m
e is the corner
where de denotes the distance from the corner feature point to the line, f i,k
e e
point in the current point cloud frame, and f u,k and f v,k are the two nearest neighbors
corner points in the local map.
The surface point residuals are calculated as follows:
s − f s ) · (( f s − f s ) × ( f s − f s ))|
|( f i,k u,m u,m v,m u,m w,m
ds = s − f s ) × ( f s − f s )|
(9)
|( f u,m v,m u,m w,m

where ds denotes the distance from the surface point to the plane, f i,k s is the surface point
s s s
in the current point cloud frame, and f u,m , f v,m s, and f w,m are the three nearest-neighbour
noncollinear points in the local map.
For semantics, the traditional approach is to reduce the error caused by mismatching
by directly identifying the points that are inconsistent with the semantic attributes of the
current point. However, in this paper, considering the possible inaccuracy of semantic
segmentation, especially at the edges, instead of adopting the constraint construction
by considering only the points with the same semantic categories [5], the residuals are
weighted by calculating the semantic weights. Specifically, semantic weights are first
obtained by calculating the probability of semantic consistency between the current point
and its five matches selected by the normal vector consistency-based feature association
method. Then, this semantic weight is utilized to weight the residuals to mitigate the
effect of outliers. The semantic label lic of the current point f i has the semantic distribution
Sensors 2024, 24, 7546 8 of 18

MSi = {lkm }kN=1 corresponding to the set of matching points MDi , and the semantic
consistency judgment is expressed as follows.
(
1 if lic = lkm
nik = (10)
0 otherwise
The weighted corner point residuals d′e and surface point residuals d′s are formulated
as follows:
(∑kN=1 nik )
d′e = de × (11)
N
(∑kN=1 nik )
d′s = ds × (12)
N
where de denotes the distance residuals of corner points , ds denotes the distance residuals of
surface points, n denotes the number of points in the matched points that are consistent with
the semantic attributes of the current point, and N denotes the number of matched points.
The introduction of semantic weights changes the objective of residual computation,
enabling the optimization to not only depend on geometry distances but also consider the
consistency of semantic attributes. This method effectively reduces the adverse effect of
geometric similar but semantic inconsistent points on the optimization and then improves
the accuracy of the solution. When semantic inconsistencies occur, the contribution of
wrong matches to the optimization objective function can be reduced by multiplying the
weights by less than one.
Ultimately, the optimal transformation matrix is solved by constructing a cost function
with the weighted residuals of corner and surface point distances and then optimizing this
cost function through the Gauss-Newton optimization method. The optimization formula
is as follows:
T ∗ = arg min{ ∑ d′e + ∑ d′s } (13)
T∗ e ∈ Fe
f i,k i
s ∈ Fs
f i,k i

where T ∗ denotes the optimal transform obtained by solving.

3.2. Semantic-Based Keyframe Selection

During SLAM, there is a wide overlap in consecutive frames, and optimizing all the
frames would lead to data redundancy and a huge computational burden. Hence, keyframe
selection is required. Heuristic methods usually select keyframes by calculating the relative
motion position. Specifically, when the pose transformation between the current frame and
the previous keyframe exceeds a certain translation or rotation threshold, the current frame
is set as a new keyframe, while other frames between the two keyframes are discarded. This
process can be described as follows: for the previous keyframe and the new point cloud
frames P1 and P2 , the pose transformations obtained through the odometry are T1 and T2 ,
and the new point cloud frames are defined as keyframes when the difference between
T1 and T2 satisfies the condition ∥t12 ∥ > τt or ∥θ12 ∥ > τr . Where τt and τr represent the
translation and rotation thresholds, ∥t12 ∥ and ∥θ12 ∥ represent the translation vector norm
and the rotation angle norm. We ensure by using distance and angle thresholds that the
system selects a new keyframe only when the pose has sufficient translation or rotation.
However, the selection of keyframes based only on pose changes may ignore frames
with drastic changes in the environment structure, resulting in the loss of keyframes. This
phenomenon occurs because drastic changes in the scene are not immediately reflected as
significant changes in the pose, resulting in the keyframe selection process not capturing
these changes in time and thus affecting the optimization results. When the scene changes
rapidly, more keyframes can help the SLAM system to adapt to these changes better and
avoid error accumulation caused by drastic changes in the scene.
In addition, in complex environments, the presence of dynamic objects increases
the uncertainty and noise in the SLAM system, causing discontinuity in the observation
data and affecting the accuracy of the pose estimation. Therefore, in the scenes with
Sensors 2024, 24, 7546 9 of 18

dynamic objects, selecting more stable keyframes can make the pose estimation process
smoother, reduce the error association, and reduce the phenomenon of pose jumps due to
dynamic objects.
Therefore, given the limitations of heuristics at the level of understanding the environ-
ment, we propose to introduce a semantic description of the scene. Usually, the distribution
of objects in the same scene region is relatively stable, and the semantic class of objects
can be used to characterize the region. Since the semantic changes between neighboring
point cloud frames can reflect the apparent differences in the environment, it enables the
SLAM system to capture the environmental changes more accurately in complex dynamic
environments and rapidly changing scenes. Therefore, we introduce semantic label infor-
mation in the keyframe selection stage to optimize the keyframe selection strategy through
semantic change detection and semantic stability detection. Compared with heuristics, our
approach better reflects the scene characteristics.
Before describing our keyframe selection method in detail, the relevant statistical
background is first provided. The KL (Kullback-Leibler) divergence is derived from the
concept of information entropy, which is used to describe the amount of information that is
available when generating a random variable from a distribution. For a random variable x
given a probability distribution P( x ), the information entropy H ( P) is defined as follows:

H ( P) = − ∑ P( x ) log P( x ) (14)
x

Based on information entropy, KL divergence is used as a measure of the relative

difference between two probability distributions. Assume that for a random variable x,
there exist two probability distributions P( x ) and Q( x ), and assume that P( x ) > 0 and
Q( x ) > 0 hold for all x. If x is a discrete random variable, define the KL dispersion from P
to Q to be:
P( x )
DKL ( P∥ Q) = ∑ P( x ) log (15)
x Q( x )
In our keyframe selection strategy, the points numbers of different semantic categories
in the two lidar frames are modeled as a discrete random variable, requiring the probability
distribution of each semantic category to be non-negative and normalized, i.e., the probabil-
ity of all categories sums up to 1, and ensuring that the probability value of each semantic
category is greater than 0 in the previous frame to ensure that the calculation of the KL
divergence is valid. First, the semantic distribution of the current frame is established by
calculating the number of points of each semantic category. We then use KL divergence to
describe the difference in semantic probability distributions between the current frame and
the previous frame.
The semantic distribution is defined as the percentage of the number of different
semantic points and is calculated as follows:
ni
Pi = (16)
N
where N is the total count of points in the point cloud and ni signifies the count of points
with semantic label i within the point cloud.The KL divergence is calculated as follows:
n Pc
DKL = ∑ Pic log( Pil ) (17)
i =1 i

where Pic and Pil are the frequency of each semantic class object in the current and previous
frames. A larger KL divergence indicates a larger semantic difference between the current
frame and the previous frame, and then the current frame is more likely to be selected as
a keyframe. Compared with keyframe selection relying only on pose changes, modeling
semantic information using KL divergence can more acutely capture changes in object
Sensors 2024, 24, 7546 10 of 18

categories and distributions in the scene, especially when the scene changes drastically,
but the pose changes are not obvious.
Secondly, we exploit semantic information to select point cloud frames that contain a
larger number of absolute static object labels. The semantic stability score SS of the current
frame is indicated by calculating the proportion of static objects relative to dynamic objects
with the following formula:
ns
SS = (18)
nd + 1
where ns is the number of absolute static objects, and nd is the number of dynamic objects.
The semantic stability score reflects the static degree of the current frame. It effectively
reduces the interference caused by dynamic objects by favoring the selection of frames with
higher static as keyframes.
Next, based on the calculated semantic distribution probability and stability score,
different semantic weights Wse are defined to adjust the threshold of keyframe selection
based on translation and rotation in the traditional geometry method. The equations for
calculating the translation and rotation thresholds are as follows:

0
 if SS > 1 ∗ tss
m= 1 if 1 ∗ tss < SS < 1.2 ∗ tss (19)

2 if SS > 1.2 ∗ tss


Wse = 0.9m × 1 − DKL (20)

τt′ = Wse × τt (21)
τr′ = Wse × τr (22)
where τt and τt′ are the original and adjusted translation thresholds; τr and τr′ are the
original and adjusted rotation thresholds, and the parameter m is related to the semantic
stability score SS. To dynamically change the selection of keyframes according to the static
degree of the scene, we set the threshold tss to 10. In addition, to balance the system
accuracy and computational efficiency and to ensure sufficient spatial and perspective
changes between keyframes, the translation distance threshold τt is set to 1.0 meters, and the
rotation angle threshold τr is set to 0.2 radians. By introducing semantic weights, keyframes
can be selected more flexibly than traditional geometric thresholds.
The use of semantic information to select keyframes provides a more accurate under-
standing of environmental changes. It improves the ability of the system to perform in
different environments compared to selecting keyframes using only pose transformations.
Finally, by judging the relationship between the interframe pose change and the ad-
justed threshold, it is decided whether the current frame is a keyframe or not: if the change
is larger than the threshold, then the frame is selected as a keyframe. Otherwise, the current
frame is discarded. The selected keyframes are not only used for factor map optimization
but are also accumulated to build local and global maps gradually. The more comprehen-
sive keyframe selection method can effectively improve the accuracy and robustness of the
position estimation.
The keyframe selection method process is shown in Algorithm 2.
Sensors 2024, 24, 7546 11 of 18

Algorithm 2 Semantic constraint-assisted pose estimation

input: current frame point cloud set Lk = { p1k , p2k , ..., pkm },previous keyframe point cloud
set Lk−1 = { p1k−1 , p2k−1 , ..., pkm−1 }, pose transformation between two frames T k−1,k
output: keyframe or not
1. for pik−1 in Lk−1 do
2. Ncar k −1 + + ; N k −1 + + ; N k −1
pole building + +
3. end for
4. for pik in Lk do
5. Ncar k + +; N k k
pole + +; Nbuilding + +
6. end for
7. DKL ← KL( Ncar k −1 , N k −1 , N k −1 k k k
pole building , Ncar , Npole , Nbuilding )
k
8. S ← score( Nbuilding k , Nk )
, Npole car
9. w ← adaption( DKL , S)
10. angleT ′ , distT ′ ← modi f y( angleT, distT, w)
11. x, y, z, roll, pitch, yaw ← getTransBetween( T k−1,k ))
12. if(check(x,y,z,roll,pitch,yaw,angleT ′ ,distT ′ ))
13. return false
14. end if
15. return true

4. Experiments and Results

To validate the advantages of our algorithm, we designed a series of experiments to
evaluate and analyze the proposed method and system. The experiments are based on
the KITTI dataset, and comparative experiments are conducted in three areas: associative
feature filtering, keyframe selection strategy, and pose estimation results. Firstly, we tested
the effectiveness of the proposed association method by comparing the number of times
nearest neighbor point usage at different distances; then, we compared the number of
point cloud frames successfully selected as keyframes with the total number of all point
cloud frames, which showed that the method was able to increase the number of effective
keyframes; lastly, we compared the performance of the SLAM system in terms of pose
estimation accuracy with other SLAM systems, to verifying the advantages of our system
in improving the accuracy. All experiments are executed on an experimental platform with
Intel i7-8700 CPU, 32 G RAM, and Ubuntu 18.04 system environment.

4.1. Datasets and Evaluation Method

In order to verify the effectiveness of the proposed algorithm, we conduct experiments
on the public dataset KITTI [21]. The KITTI dataset provides LiDAR data, image data,
IMU/GPS data, and calibration information covering 22 sequences. The data were collected
around the medium-sized city of Karlsruhe, in rural areas and on motorways, and the scenes
contain objects such as vehicles, pedestrians, and buildings. We validated the experiments
using 10 sequences with ground truth in the KITTI dataset. Table 1 lists information about
the KITTI sequences used in the experiments. These sequences cover a variety of scenes,
such as urban, rural and high-speed roads, and the length of the trajectories ranges from
394 m to 5067 m, including trajectories with and without loops. Therefore, the performance
of the proposed algorithm can be fully verified.
To quantitatively evaluate the experimental results, we use Absolute Trajectory Error
(ATE) as a metric to assess the accuracy of the pose estimation. ATE measures the accuracy
of the system on the overall trajectory by calculating the absolute Root Mean Square Error
(RMSE) between the estimated pose and the true pose for each keyframe.
Sensors 2024, 24, 7546 12 of 18

Table 1. Scene information of each sequence in KITTI dataset.

Sequence Number of Scans Trajectory Length (m)

01 1101 2453
02 4661 5067
04 271 393
05 2761 2205
06 1101 1232
07 1101 694
09 1591 1705
10 1201 919

4.2. Feature Association Comparison

To verify the performance of the feature association method, we tested the sequences 01
and 06 in KITTI dataset, respectively. In the experiments, the effectiveness of the traditional
association method and the proposed method is compared by analyzing the number of
times association use at different distances from the nearest neighbors.
In traditional methods, the five nearest neighbors are usually directly selected as
association points to construct the residuals. Our proposed method introduces normal
vector consistency to filter five points from 15 nearest neighbors as association points. To vi-
sualize the effectiveness of the method, we sort the candidate matching points according
to the nearest-neighbor distance and count the number of times the serial numbers of the
matching points are finally used for feature association. The formula for calculating the
total number of times Ni that a point with matching point serial number i is selected in the
feature association process is shown below:

k
Ni = ∑ nj (23)
j =1

where n j denotes the number of times the point with matching point number i is selected
in the jth feature association process,k denotes the total number of times of feature associa-
tion process.
Figures 2 and 3 show the statistics of the number of times the match point sequence
numbers in sequences 01 and 06, respectively. Figures 2a,b and 3a,b indicate the statistics
of the traditional and proposed methods, respectively. The horizontal axis of the figure
indicates the serial number of the candidate match point and the vertical axis shows the
number of times the match point is selected as a match point in the feature association
phase. It can be observed that in the traditional method, only the five nearest neighbor
points are selected as matching points. Whereas the proposed algorithm is not limited to
the first five nearest neighbor points, the 6th to 15th nearest neighbor points are also used
as matching points. Although these matching points are used less frequently than the first
5 points, they still account for a larger proportion of feature association.
To further validate the effectiveness of the method, we conducted comparative exper-
iments on pose estimation accuracy for sequences 01 and 06. The results show that the
accuracy of the proposed method is improved by 2.77% and 2.86% in these two sequences,
respectively. Therefore, by introducing normal vector information, the associated points
can be selected more accurately, thus improving the accuracy of feature point matching.
Sensors 2024, 24, 7546 13 of 18

(a) (b)

Figure 2. The number of times use of 01 sequence correlation points. (a) Traditional method.
(b) Proposed method.

(a) (b)

Figure 3. The number of times of 06 sequence correlation points. (a) Traditional method. (b) Proposed
method.

4.3. Keyframe Selection Strategy Comparison

To verify the effectiveness of the proposed keyframe selection method, we analyzed
sequences 06 and 07 of the KITTI dataset. Sequence 06 has a trajectory length of 1232 m
and contains a total of 1101 frames of lidar scan data. Sequence 07 has a trajectory length of
694 m and also contains 1101 frames of LiDAR scan data. The traditional method selects
keyframes based on the geometric difference thresholds of translation and rotation between
neighboring frames. In contrast, the proposed method evaluates the inter-frame similarity
and stability by combining the semantic information to adjust the geometric thresholds
for keyframe selection dynamically. Our experiments compare the traditional keyframe
selection method with the proposed semantic-based improved method.
The experimental results are shown in Table 2. In sequence 06, the proposed method
selects 545 keyframes, while the traditional method selects 534 keyframes, and the attitude
estimation accuracy is improved by 3.64%. In sequence 07, the proposed method selects
552 keyframes, while the conventional method selects only 424 keyframes, and the pose
estimation accuracy is improved by 0.38%. It can be seen that the proposed keyframe
selection strategy selects significantly more keyframes than the traditional method, which
indicates that the proposed approach improves the estimated pose accuracy by selecting
more keyframes.

Table 2. Comparison of keyframe selection methods.

Number of Key Number of Key

Sequence Number of Frames Frames Frames ATE/m
(Traditional Method) (Proposed Method)
06 1101 532 545 14.001
07 1101 424 552 0.437
Sensors 2024, 24, 7546 14 of 18

To visualize the distribution of keyframes in the SLAM process more intuitively, we

counted the distribution of keyframe ordinal numbers of the two methods in the trajectory.
As shown in Figure 4, red identifies the traditional method and blue identifies the improved
method. The X-axis indicates the frame serial number, and the Y-axis indicates the number
of times the keyframes are selected.
For the proposed keyframe selection method, in Sequence 06, keyframes are selected
more frequently between frames 200 to 300 and 600 to 700, where the robot is performing
a turning operation and the viewpoint undergoes a continuous transformation in these
regions. Similarly, Sequence 07 has significantly more frequent keyframe selection between
frames 300 to 400 and in the range of 500 to 700 frames, and there are continuous turns in
these regions. These observations show that the proposed method exhibits more sensitive
and accurate key frame selection in regions with changing environmental characteristics,
effectively improving the accuracy of system attitude estimation.

(a)

(b)

Figure 4. In the same sequence, the keyframe sequence numbers selected by the traditional pose
change method and the method using the scene semantic information were compared. (a) sequence
06. (b) sequence 07.

4.4. Pose Estimation Comparison

To assess the precision and stability of the proposed LIO-SAM++, we contrast its
performance with classical SLAM, and the comparison model is as follows:
(1) LEGO-LOAM [1] filters noise through point cloud segmentation, matches lidar point
cloud frames using surface and corner features, and reduces cumulative drift errors
using the loop closure detection module, thus enhancing the accuracy of overall
localization accuracy.
(2) LIO-SAM [2] performs pose estimation by feature matching between point cloud
frames and local maps and constructs a factor graph for back-end optimization.
(3) FAST-LIO [3] gives iterative Kalman filtering to estimate the pose transform and
updates the feature point map based on the state estimation results.
(4) SUMA++ [4] introduces semantic information based on the SUMA, filters the dynamic
targets through the semantic difference between the observation at the current moment
and the previous observation, and introduces semantic constraints into the residuals,
thus generating semantic maps and improving localization accuracy in complex envi-
ronments. Notably, SUMA++ also integrates RangeNet++ for semantic segmentation.
Sensors 2024, 24, 7546 15 of 18

We performed a series of experiments across various sequences of the KITTI dataset

to assess the localization accuracy of diverse SLAM algorithms. Table 3 summarises the
absolute trajectory errors of different SLAM algorithms in eight scenes. The experimental
outcomes indicate that the proposed LIO-SAM++ has an average absolute trajectory error
of 6.56m over multiple sequences, which outperforms other algorithms overall. Compared
to LEGO-LOAM, LIO-SAM, FAST-LIO, and SUAM++, the improvements are 13.1%, 8.6%,
15.2%, and 19.4%, indicating that the proposed algorithm has higher localization accuracy
in most scenes. It is worth noting that LEGO-LOAM shows significant drift or even fails
in some cases in many sequences, which is mainly due to the point cloud segmentation
module affecting the number of extracted features, resulting in the inability to associate
features accurately. In sequences 05 and 07, due to the presence of a large number of
dynamic objects, the trajectories generated by the LIO-SAM++ drifted to some extent,
mainly due to incorrectly associated feature points generated by the dynamic objects.
Nevertheless, LIO-SAM++ still exhibits strong trajectory consistency overall, indicating
that it still has high robustness and stability when dealing with diverse scenes.

Table 3. Comparison of absolute trajectory error on the KITTI datasets (RMSE of ATE [m]).

LIO-SAM++
Sequence LEGO-LOAM LIO-SAM FAST-LIO SUMA++
(Ours)
01 - 20.31 14.96 - 19.76
02 - 9.21 18.87 30.25 6.25
04 0.39 0.25 0.16 0.94 0.25
05 3.29 0.81 3.14 1.11 0.87
06 26.84 14.51 4.90 1.14 13.99
07 2.02 0.43 0.84 0.93 0.44
09 11.76 9.73 14.08 14.65 8.95
10 2.12 2.17 4.94 8.81 1.97
Average 7.55 7.18 7.74 8.14 6.56

For a more in-depth analysis, we chose sequences 02, 09 and 10 from the KITTI dataset
for trajectory comparison, which cover diverse scenarios such as urban and rural areas. We
show the trajectory comparisons of different algorithms and the deviation from the ground
truth, respectively.
Figure 5 depict the comparison results of trajectories, relative position, and relative
rotations for each algorithm in sequence 02.Sequence 02 is a 5067-m-long urban scene
sequence containing a large number of curves. In this scene, the trajectories generated
by the LIO-SAM++ algorithm are very close to the ground truth, especially in the curved
sections, where it significantly outperforms other algorithms.In comparison, the trajectories
of the other algorithms and the deviations of other algorithms are larger. That is because
the keyframe thresholds set by LIO-SAM++ through the adaptive understanding of the en-
vironment can better adapt to the scene changes, avoiding the omission of valid keyframes
when traveling in the curves and thus enabling more accurate estimation of the trajectory.
In Sequence 09, the vehicle navigates through a rural road environment, and there are
many objects like vegetation, buildings, and roads, as well as significant elevation changes
along the route. Figure 6 displays the trajectory comparison outcomes for this sequence,
where the FAST-LIO has significant trajectory errors.In contrast, the trajectories obtained
by LEGO-LOAM and LIO-SAM are largely coincident, and the LIO-SAM++ algorithm
behaves closer to the true values.
Sequence 10 is also in complex rural road conditions and contains a large number of
dynamic objects. Figure 7 illustrates that LIO-SAM++ surpasses other methods, with its
trajectory almost exactly aligning with the real trajectories, especially in the error control
of translational and rotational poses. This is due to the more stable and accurate features
provided by the proposed feature association method, which ensures the robustness of the
Sensors 2024, 24, 7546 16 of 18

pose estimation through normal vector and semantic consistency constraints, thus reducing
the system drift.
In addition, from the intermediate subplots (labeled ‘b’) of Figures 5–7, it can be seen
that the different algorithms have more consistent trajectory estimation in the x-axis and
z-axis directions. Still, there is a significant difference in the y-axis direction. This is due to
the lower measurement resolution of LiDAR in the y-axis direction and less observation
information from the ground, resulting in insufficient y-axis constraints. Nevertheless,
the proposed LIO-SAM++ still shows strong robustness in the y-axis.
Overall, the experiment results verify the effectiveness of the proposed algorithm in
various scenes and demonstrate that it can improve the accuracy of the SLAM system.

(a) (b) (c)

Figure 5. Comparative analysis of the estimated trajectories against the GT in sequence 02.
(a) trajectories. (b) position. (c) rotation.

(a) (b) (c)

Figure 6. Comparative analysis of the estimated trajectories against the GT in sequence 09.
(a) trajectories. (b) position. (c) rotation.
Sensors 2024, 24, 7546 17 of 18

(a) (b) (c)

Figure 7. Comparative analysis of the estimated trajectories against the GT in sequence 10.
(a) trajectories. (b) position. (c) rotation.

5. Conclusions
In this paper, we propose a lidar-inertial semantic SLAM algorithm with association
optimization and keyframe selection. The algorithm obtains more accurate feature associa-
tions by geometrically modeling the local point cloud and filtering out associated points
that are not in the same plane. Then, semantic consistency is added to the estimation
problem using the semantic label as a constraint weighting term, and this approach takes
semantic inaccuracies fully into account. In the optimization process, the distribution of
semantic labels is used to calculate the inter-frame differences and the degree of stability of
the current frame to focus on the changing characteristics of the scene, thus achieving a
more reasonable keyframe selection.
Experimental validation on the KITTI Odometry dataset demonstrates that our al-
gorithm enhances the precision of pose estimation and simultaneously generates clearly
structured semantic point cloud maps. Even though the proposed framework provides
reliable results, the work of semantic mapping still faces some challenges. Currently,
the only open-source dataset with semantic labeling of LiDAR frames is SemannticKITTI,
which limits the generality of the point cloud segmentation network used. In addition,
the adaptability in dynamic environments has not been fully validated. In subsequent
research, we aim to explore the potential of semantic information further to handle dynamic
objects better and, consequently, to deal with more complex real-world situations.

Author Contributions: Conceptualization, B.S. and W.X.; methodology, B.S.; software, B.S.; valida-
tion, B.S.; formal analysis, B.S., Z.G. and X.Q.; investigation, B.S. and Z.G.; resources, W.X.; data
curation, B.S.; writing—original draft preparation, B.S.; writing—review and editing, B.S. and W.X.;
visualization, B.S.; supervision, X.P. All authors have read and agreed to the published version of the
manuscript.
Funding: This research was funded by Cultivation Fund of Chinese Academy of Sciences (no.
KGFZD-145-24-33).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The dataset used in the paper is the public KITTI Odometry, which
can be downloaded at: https://www.cvlibs.net/datasets/kitti/eval_odometry.php (accessed on 24
November 2024).
Conflicts of Interest: The authors declare no conflict of interest.
Sensors 2024, 24, 7546 18 of 18

References
1. Shan, T.; Englot, B. LeGO-LOAM: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain.
In Proceedings of the 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Madrid, Spain,
1–5 October 2018; pp. 4758–4765.
2. Shan, T.; Englot, B.; Meyers, D.; Wang, W.; Ratti, C.; Rus, D. LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing
and Mapping. In Proceedings of the 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Las Vegas,
NV, USA, 25–29 October 2020; pp. 5135–5142.
3. Xu, W.; Zhang, F. FAST-LIO: A fast robust LiDAR-inertial odometry package by tightly-coupled iterated Kalman filter. IEEE
Robot. Autom. Lett. 2021, 6, 3317–3324. [CrossRef]
4. Chen, X.; Milioto, A.; Palazzolo, E.; Giguère, P.; Behley, J.; Stachniss, C. SuMa++: Efficient LiDAR-based Semantic SLAM.
In Proceedings of the 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Macau, China,
4–8 November 2019; pp. 4530–4537.
5. Li, L.; Kong, X.; Zhao, X.; Li, W.; Wen, F.; Zhang, H.; Liu, Y. SA-LOAM: Semantic-aided LiDAR SLAM with Loop Closure. In
Proceedings of the 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi’an, China, 30 May–5 June 2021;
pp. 7627–7634.
6. Ji, H.; Singh, S. LOAM: Lidar Odometry and Mapping in Real-time. Robot. Sci. Syst. 2014, 2, 9.
7. Guo, S.; Rong, Z.; Wang, S.; Wu, Y. A LiDAR SLAM With PCA-Based Feature Extraction and Two-Stage Matching. IEEE Trans.
Instrum. Meas. 2022, 71, 8501711. [CrossRef]
8. Pan, Y.; Xiao, P.; He, Y.; Shao, Z.; Li, Z. MULLS: Versatile LiDAR SLAM via Multi-metric Linear Least Square. In Proceedings of
the 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi’an, China, 30 May–5 June 2021 ; pp. 11633–11640.
9. Xu, M.; Lin, S.; Wang, J.; Chen, Z. A LiDAR SLAM System With Geometry Feature Group-Based Stable Feature Selection and
Three-Stage Loop Closure Optimization. IEEE Trans. Instrum. Meas. 2023, 72, 8504810. [CrossRef]
10. Zhao, Z.; Zhang, W.; Gu, J. Lidar mapping optimization based on lightweight semantic segmentation. IEEE Trans. Intell. Veh.
2019, 4, 353–362. [CrossRef]
11. Du, S.; Li, Y.; Li, X.; Wu, M. LiDAR Odometry and Mapping Based on Semantic Information for Outdoor Environment. Remote
Sens. 2021, 13, 2864. [CrossRef]
12. Forster, C.; Pizzoli, M.; Scaramuzza, D. SVO: Fast semi-direct monocular visual odometry. In Proceedings of the 2014 IEEE
International Conference on Robotics and Automation (ICRA), Hong Kong, China, 31 May–7 June 2014; pp. 15–22.
13. Mur-Artal, R.; Montiel, J.M.M.; Tardós, J.D. ORB-SLAM: A Versatile and Accurate Monocular SLAM System. IEEE Trans. Robot.
2015, 31, 1147–1163. [CrossRef]
14. Engel, J.; Koltun, V.; Cremers, D. Direct Sparse Odometry. IEEE Trans. Pattern Anal. Mach. Intell. 2018, 40, 611–625. [CrossRef]
[PubMed]
15. Qin, T.; Li, P.; Shen, S. VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator. IEEE Trans. Robot. 2018, 34,
1004–1020. [CrossRef]
16. Kuo, J.; Muglikar, M.; Zhang, Z.; Scaramuzza, D. Redesigning SLAM for Arbitrary Multi-Camera Systems. In Proceedings of the
2020 IEEE International Conference on Robotics and Automation (ICRA), Paris, France, 31 May–31 August 2020.
17. Jiao, J.; Zhu, H.Y.Y.; Liu, M. Robust Odometry and Mapping for Multi-LiDAR Systems with Online Extrinsic Calibration. IEEE
Trans. Robot. 2022, 38, 351–371. [CrossRef]
18. Lin, Y.; Dong, H.; Ye, W.; Dong, X.; Xu, S. InfoLa-SLAM: Efficient Lidar-Based Lightweight Simultaneous Localization and
Mapping with Information-Based Keyframe Selection and Landmarks Assisted Relocalization. Remote Sens. 2023, 15, 4627.
[CrossRef]
19. Milioto, A.; Vizzo, I.; Behley, J.; Stachniss, C. RangeNet ++: Fast and Accurate LiDAR Semantic Segmentation. In Proceedings
of the 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Macau, China, 4–8 November 2019;
pp. 4213–4220.
20. Behley, J.; Garbade, M.; Milioto, A.; Quenzel, J.; Behnke, S.; Stachniss, C. SemanticKITTI: A Dataset for Semantic Scene
Understanding of LiDAR Sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV),
Seoul, Republic of Korea, 27 October–2 November 2019.
21. Geiger, A.; Lenz, P.; Urtasun, R. Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Providence, RI, USA, 16–21 June 2012; pp. 3354–3361.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.