2023 IEEE International Conference on Robotics and Automation (ICRA)

Workshop on Active Methods in Autonomous Navigation

29 May – 2 June, 2023, London, UK

Visual-based and Lidar-based SLAM Study for Outdoor Environment

Siti Sofiah Binti Mohd Radzi 1, Siti Sarah Binti Md. Sallah 1, Nazlee Azmeer bin Massuan 1, Lee Ming
Yi 1, Qamarul Aiman Bin Tajul Ariffin 1, Bayhaqi Bin Mohd Jailani1, and Hon Hock Woon 1

Abstract— Malaysia's palm oil plantation industry is labor making. It is divided into two main types [3]: visual SLAM
intensive. It is estimated that there were 505,972 workers in 2012 and Lidar-Based SLAM. Visual SLAM, or VSLAM, uses
and this number consists mainly of foreigners of about 76.5% cameras and image sensors to capture data, offering
and local 23.5%. To address the labor-intensive nature of the advantages such as rich information, affordability, lightweight
industry and reduce costs, mechanization and automation have design, and small size. However, it is sensitive to lighting
emerged as potential solutions. However, the dense foliage of conditions. On the other hand, Lidar-Based SLAM primarily
matured oil palm trees, along with their overlap with
neighboring trees, obstructs crucial information beneath the
relies on laser or distance sensors, providing greater precision
canopy. Aerial mapping or Google map images cannot reliably and being unaffected by lighting conditions. Nevertheless,
provide the necessary information for unmanned ground high-resolution lidar sensors can be expensive. Both types of
vehicles (UGVs) to navigate the under-canopy environment. SLAM contribute to navigation and decision-making
Furthermore, GPS signals may be degraded or inaccessible processes by enabling simultaneous localization and mapping
under the canopy. Consequently, researchers have developed in various applications.
Vision-based and Lidar-based navigation methods, specifically The main objective of a 3D SLAM system is to determine
tailored for GPS-denied environments, such as Simultaneous the robot's position and orientation while creating a map of the
Localization and Mapping (SLAM). This survey paper aims to environment using sensor data. To build the map, classic
explore the current research on Visual-based and Lidar-based
navigation in outdoor environment without relying on GPS
SLAM techniques use front-end odometry, which relates
sensor. Experiments are carried out in outdoor environment to subsequent sensor scans, such as lidar or stereo camera point
simulate situations without GPS coverage. The resulting maps clouds. The typical method for relating these scans is iterative
generated from these methods were then qualitatively analyzed. closest point (ICP), but it can be computationally expensive
when dealing with a large number of points. Another approach
is feature-matching, which reduces computation costs.
I. INTRODUCTION
However, both methods still result in some accumulated error
The robotics community has witnessed a significant rise in over time, leading to inaccurate odometry. To address this,
outdoor applications across various sectors such as back-end optimization is used. There are two types of back-
agriculture, search and rescue, mining, defense, environment end optimization methods: filtering approaches like Extended
monitoring, and planetary exploration. In the agricultural Kalman Filter (EKF), Unscented Kalman Filter (UKF), or
sector, where productivity relies on efficiency, reliability, and particle filter, which correct the robot's state in real-time as
precision, reducing the need for manual labor is crucial as it new measurements are available, and graph optimization
constitutes a major portion of field operation costs. For approaches, which optimize the full robot trajectories using a
example, Malaysia's labor-intensive oil palm plantation complete set of measurements to enhance odometry accuracy.
industry employed approximately 505,972 workers in 2012, The methods used in graph optimization are for example such
with foreigners accounting for 76.5% and locals 23.5%.[1] as iSAM (Incremental Smoothing And Mapping) [4],
Mechanization and automation are crucial for reducing labor GTSAM (Georgia Tech Smoothing and Mapping) [5], and
costs and maintaining competitiveness in agriculture. G2O(General Graph Optimization) [6]. In SLAM algorithm,
However, the agricultural environment is challenging, with the loop closure detection is added to detect whether the path
dynamic and unstructured conditions. Accurate mapping and has formed a loopback. This is because the current trajectory
localization are necessary for autonomous navigation in such accumulates error by time and form a significant trajectory
environments. Traditional GPS signals may be limited or drift. Loop closure detection eliminates drifts by eliminating
unavailable, requiring the use of Simultaneous Localization visited locations to create a more accurate map. During each
and Mapping (SLAM) techniques. SLAM compensates for mapping session, global loop closure detection is used to
GPS limitations by utilizing sensors like stereoscopic cameras figure out when the robot goes back to a previous map[7].
or Lidar, which are combined with secondary sensors like Global loop closure is important because it can fix errors that
IMUs, wheel odometry, or GPS to maintain the 3D spatial build up as the robot moves through the environment. Local
structure of SLAM. The combination of multiple sensor inputs loop closure, on the other hand, can make the map more
enables vehicle path planning, obstacle avoidance, and consistent and complete in a given area. A user may not revisit
manipulation. [2] to a particular point in the scene, which makes global loop
closure impossible. Therefore, local loop closures are used to
II. SLAM ALGORITHMS refine the initial pose estimate [8]. There are several SLAM
SLAM (Simultaneous Localization and Mapping) plays a methods, for example, RTABMAP [9], ORB-SLAM3 [10],
significant role in autonomous vehicles and robot decision LEGO-LOAM [11], LIO-SAM [12, 13], LVI-SAM [14] and

*Research is supported by Malaysian Ministry of Science, Technology [email protected],[email protected],aiman.ariffin@mimos.

and Innovation (MOSTI) under the strategic research grant (SRF). 1 Authors my, [email protected], [email protected])
are with MIMOS Berhad Malaysia under Advanced Intelligence Lab.
([email protected], [email protected],
others. In this study, we have compared two SLAM roughness of points over a local region. Instead of using every
(Simultaneous Localization and Mapping) algorithms: lidar for computing and adding factors, the LIO SAM adopt the
RTABMAP and LIO-SAM using a combination of 3D Lidar, keyframe selection. The factor graph is optimized upon the
stereo camera, and IMU (Inertial Measurement Unit) as insertion of a new node using incremental smoothing and
inputs. RTABMAP has the capability to use either Lidar or mapping with the Bayes tree (iSAM2) [17].Warku et. al [13]
visual camera input for its odometry node, whereas LIO-SAM use a Velodyne (VLP-16) and Xsens MTI-G-700 to obtain data
specifically utilizes Lidar and IMU data. For this research, we from inertial measurement devices to create a 3D mapping of
aimed to investigate and compare the qualitative differences the indoor and outdoor environment and visualize the created
map using the LIO-SAM method. The map optimization
in map accuracy between these two algorithms, despite both
optimizes lidar odometry factor and GPS factor. This factor
of them incorporating Lidar as one of the input sensors.
graph is maintained consistently throughout the whole test. The
A. RTABMAP IMU pre-integration optimizes IMU and the lidar odometry
RTAB-MAP (Real-Time Appearance-Based Mapping) [9, factor and estimates the IMU bias. This factor graph is
periodically reset and guarantees real-time odometry
15] is a graph-based SLAM approach widely used in mobile
estimation with frequency IMU. Fig. 1 illustrates the factor-
robot navigation within the Robot Operating System (ROS). It graph structure in LIO-SAM flow diagram. Loop closure
supports various data types such as RGB-D, stereo, and detected by Euclidean distance radius search and ICP
LiDAR, and is versatile in handling mixed modalities. RTAB- registration significantly degrades real-time performance [18].
MAP can work with different odometry approaches, including
visual, LiDAR, or wheel-based methods. For visual odometry,
it uses Frame-To-Map (F2M) or Frame-To-Frame (F2F)
techniques, while LiDAR odometry involves Scan-to-Map
(S2M) or Scan-to-Scan (S2S) methods. The latter processes
3D point clouds from LiDAR scans, down-samples them,
calculates normals, and applies the Iterative Closest Point
(ICP) algorithm to estimate the transformation between fixed
and moving point clouds. RTAB-MAP incorporates an Figure 1 Factor-graph structure of LIO-SAM[12]
appearance-based loop closure detector using a bag-of-words
approach, which helps identify whether a new image
III. EXPERIMENTAL SETUP AND RESULTS
corresponds to a previous or new location[15].When a loop
closure hypothesis is accepted, a new constraint is added to
A. Experimental Setup
the graph of the map, whereupon a graph optimizer minimizes
the errors in the map. When a loop closure occurs, the global Our experimental site is located at the MIMOS campus in
map should be re-assembled according to all new optimized Kuala Lumpur, Malaysia. The total size of the MIMOS site is
poses for all nodes in the map's graph. For loop closure approximately 133288 m2. The custom hardware setup with
detection, the current bag-of-words approach is dependent on Scooter-based is shown in Figure 2(a). A ROS melodic
a camera, meaning that a camera is always required even if operating system and equipped with sensors such as 3D Lidar
VLP-16, Xsens IMU 710 with external receiver and a unit
LIDAR is used for odometry.
depth camera ZED 2i. The speed of the scooter while recording
is 12km/h. The analysis of the 3D point cloud produced from
B. LIO SAM MIMOS Berhad area is divided into three subareas; A1, A2 and
LIDAR is commonly utilized alongside other sensors, such A3 as shown in Fig. 2(b). A1 has more features, mainly parking
as IMU and GPS, to estimate the state and create map. There lots. A2 contains more challenging roads (elevations, slopes,
are two primary design approaches for Lidar mapping that and bumps), and A3 is mostly consists of unoccupied parking
involve sensor fusion: loosely-coupled fusion and tightly- lots. The user drives the scooter around MIMOS Berhad in
coupled fusion. In the tightly-coupled approach, the pre-
integrated IMU measurements are typically leveraged to
rectify the skewed point clouds. Lidar Inertial Odometry, a
tightly-coupled method, employs all IMU data directly and
optionally incorporates a global positioning system (GPS) to
correct height errors. To achieve an accurate map over large
outdoor areas, sensor fusion between LIDAR/IMU and a high-
precision GPS sensor is essential.[16]. LIO-SAM formulates
lidar inertial odometry on a factor graph suitable for multi-
sensor fusion and global optimization [12]. In LIO-SAM four
types of factors are:(a) IMU pre-integration factors, (b) lidar
odometry factors, (c) GPS factors, and (d) loop closure factors. (a)
It estimates states utilizing tightly coupled IMU integration and (b)
Lidar odometry via factor graph optimization. In Lidar Figure 2. (a) Experimental setup for hand-held devices based on
odometry factor, a feature extraction is performed on new lidar motorized scooter. (b) Top view of the image captured by Google Earth.
scan. Edge and planar features are extracted by evaluating the The datasets are collected in 3 different subareas of MIMOS: A1, A2, A3
order to collect the dataset, then pivots back to the starting identify and rectify loop closures in order to enhance map
locations. accuracy and maintain consistent localization throughout the
mapping process.
B. Results Fig. 3(a) shows the top view of generate map of A1 using
For qualitative evaluation, RTABMAP and LIO-SAM
were utilized to generate a comprehensive 3D map of the
environment. To evaluate loop closure detection, markers
were placed in specific areas and compared using both
methods. In cases where a false loop closure was identified, (a)
it was classified as a false positive, indicating incorrect loop
closure detection. Conversely, if no loop closure was
detected in a similar previously visited location, it was
considered a false negative, indicating a missed loop
(b)

(c)

Figure 4 . (a) RTABMAP for A1: Local loop closure detected due to
identical objects in red box. The yellow dotted are inliers found more
than one time in the image. It is same objects’ features (drain) found
but at different location (b) RTABMAP for A2: No global closed loop
(a) (b) (c) is detected, even if the start and end points are at the same location.
This is using the default parameters (c) RTABMAP for A3:No loop
close detected

RTABMAP, Fig. 3(b) shows the trajectory of A1 with

RTABMAP, while Fig. 3(c) shows the trajectory of A1 with
LIO-SAM. Based on the observations made at marker 1, the
result indicates the presence of false positive local loop
closure detection in the correct area. However, in Fig. 4(a),
the RTABMAP database viewer shows loop closure
(d) (e) (f) detection between identical objects (specifically, a drain) but
actually at a different location. At marker 2, there is a false
negative detection where the map fails to close a loop due to
insufficient inliers in the camera image or issues with image
illumination. Similarly, at marker 3, there is an incorrect local
loop closure as the color of the identical object differs.
Moreover, at marker 4, the start and end images in map A1
7 (a), do not achieve global loop closure because of a lack of
(b) information about the U-turn, resulting in map duplication.
Fig. 3(d) provides a top view map of A2 using RTABMAP
(g) (h) (i) while Fig. 3(e) shows A2 with trajectory of RTABMAP and
Fig. 3(f) illustrate the mapping trajectory of LIO SAM in A2.
Figure 3 (a) Top view of A1 RTABMAP 3D mapping (b)RTABMAP
Mapping trajectory of A1 (c) LIO SAM Mapping trajectory of AI (d) Top Marker 5 represents a local loop closure, indicating a
view of A2 RTABMAP 3D mapping (e)RTABMAP Mapping trajectory successful detection. On the other hand, marker 6(a) indicates
of A2 (f) LIO SAM Mapping trajectory of A2 (g) Top view of A3 the absence of loop closure, which is considered a false
RTABMAP 3D mapping (h)RTABMAP Mapping trajectory of A3 (i) negative detection. The first and last camera photos displayed
LIO SAM Mapping trajectory of A3
in the RTABMAP data viewer (Fig. 4(b)) at marker 6(a)
closure.Loop closure detection can be performed through two
location show the same region and features as the vehicle
methods: local loop closure and global loop closure. Local
moves back and forth. The false negative loop closure
loop closures are focused on detecting and correcting loop
detection could be attributed to the camera's field of view. In
closures within a restricted spatial and temporal range,
contrast, using LIO-SAM, marker 6(b) demonstrates a
typically centered around the robot's current pose or
successful global loop closure detection (true positive).
trajectory segment. On the other hand, global loop closures
Fig. 3(g) shows the top view of A3, while Fig. 3(h) and
encompass the detection and correction of loop closures over
Fig. 15(i) show the mapping trajectory of A3, for RTABMAP
the entire trajectory or a significant portion of it, aiming to
and LIO SAM respectively. From the path in Fig. 3(h) and
achieve alignment and optimization at a global scale. By
Fig 3(i), no global close loops occur with RTABMAP and
employing these methods, RTABMAP and LIO-SAM aim to
LIO-SAM. Marker 7(a) indicates that there is no global loop
 because the last frames do not receive enough inliers to form (MOSTI) under the strategic research grant.
a closed loop with the starting frame, as shown in Fig. 4(c).
Fig.5(a) and Fig. 5(b) provides a close-up view of marker REFERENCES
6(b) before and after applying the LIO-SAM calculation, [1] A. Ismail, "The Effect of Labour Shortage in the Supply and Demand
where the purple line indicates the loop constraint that of Palm Oil in Malaysia," Oil Palm Industry Economic Journal, vol.
effectively closes the loop. However, there are some artifacts 13, no. 2, pp. 1-26, 2013.
leading to duplicate features in the same location, resulting [2] R. Radmanesh, Z. Wang, V. S. Chipade, G. Tsechpenakis, and D.
Panagou, "LIV-LAM: LiDAR and visual localization and mapping," in
in an inaccurate map, as shown in Fig. 5(c). For marker 7(b), 2020 American Control Conference (ACC), 2020: IEEE, pp. 659-664.
Fig. 5(d) shows the close-up of the global map. It also shows [3] G. Bresson, Z. Alsayed, L. Yu, and S. Glaser, "Simultaneous
that no global loop closure detected based on the output point localization and mapping: A survey of current trends in autonomous
cloud. The point cloud of the mapping trajectory does not driving," IEEE Transactions on Intelligent Vehicles, vo l. 2, no. 3,
overlap with the original point cloud resulting in inaccurate pp. 194-220, 2017.
[4] M. Kaess, A. Ranganathan, and F. Dellaert, "iSAM: Incremental
representation of the map. smoothing and mapping," IEEE Transactions on Robotics, vol. 24, no.
6, pp. 1365-1378, 2008.
[5] F. Dellaert, "Factor graphs and GTSAM: A hands-on introduction,"
Georgia Institute of Technology, 2012.
[6] R. Kümmerle, G. Grisetti, H. Strasdat, K. Konolige, and W. Burgard,
"g 2 o: A general framework for graph optimization," in 2011 IEEE
International Conference on Robotics and Automation, 2011: IEEE,
pp. 3607-3613.
(a) (b) [7] M. Labbe and F. Michaud, "Online global loop closure detection for
large-scale multi-session graph-based SLAM," in 2014 IEEE/RSJ
International Conference on Intelligent Robots and Systems, 2014:
IEEE, pp. 2661-2666.
[8] S. Patra, H. Aggarwal, H. Arora, S. Banerjee, and C. Arora,
"Computing egomotion with local loop closures for egocentric videos,"
in 2017 IEEE Winter Conference on Applications of Computer Vision
(WACV), 2017: IEEE, pp. 454-463.
[9] M. Labbe, "Simultaneous Localization and Mapping (SLAM) with
(c) (d) RTAB-Map," Département de Génie Électrique et Génie Informatique,
Figure 5(a) Before and (b) after loop close in A3 for marker 6(b) for 2015.
LIO SAM. (c) LIO-SAM: Artifacts presence in the map (d) LIO - [10] C. Campos, R. Elvira, J. J. G. Rodríguez, J. M. Montiel, and J. D.
SAM: Close up view from marker 7(b) Tardós, "ORB-SLAM3: An accurate open-source library for visual,"
Visual-Inertial and Multi-Map SLAM, 2020.
[11] T. Shan and B. Englot, "Lego-loam: Lightweight and ground-
IV. CONCLUSIONS AND DISCUSSION optimized lidar odometry and mapping on variable terrain," in 2018
IEEE/RSJ International Conference on Intelligent Robots and Systems
Both LIO-SAM and RTAB-Map use Lidar scans to create (IROS), 2018: IEEE, pp. 4758-4765.
3D maps, but their approaches to loop closure detection differ. [12] T. Shan, B. Englot, D. Meyers, W. Wang, C. Ratti, and D. Rus, "Lio-
LIO-SAM primarily relies on Lidar scans for odometry and sam: Tightly-coupled lidar inertial odometry via smoothing and
loop closure detection. In contrast, RTAB-Map uses Lidar mapping," in 2020 IEEE/RSJ international conference on intelligent
robots and systems (IROS), 2020: IEEE, pp. 5135-5142.
data for odometry but relies on visual information from [13] H. T. Warku, N. Y. Ko, H. G. Yeom, and W. Choi, "Three-
camera images for loop closure. RTAB-Map compares visual Dimensional Mapping of Indoor and Outdoor Environment Using
features to identify loop closures, which can be limited by LIO-SAM," in 2021 21st International Conference on Control,
factors like field of view, lighting variations, and object Automation and Systems (ICCAS), 2021: IEEE, pp. 1455-1458.
[14] T. Shan, B. Englot, C. Ratti, and D. Rus, "Lvi-sam: Tightly-coupled
similarities, leading to potential detection errors. LIO-SAM lidar-visual-inertial odometry via smoothing and mapping," in 2021
utilizes raw Lidar scans for odometry, eliminates noise, and IEEE international conference on robotics and automation (ICRA),
extracts key points for loop closure using the Iterative Closest 2021: IEEE, pp. 5692-5698.
Point algorithm. It employs a radius-search method to align [15] K. J. de Jesus, H. J. Kobs, A. R. Cukla, M. A. d. S. L. Cuadros, and D.
F. T. Gamarra, "Comparison of visual SLAM algorithms ORB-
scans from different parts of the trajectory, correcting for drift SLAM2, RTAB-Map and SPTAM in internal and external
for loop closure detection. LIO-SAM employs a tightly environments with ROS," in 2021 Latin American Robotics
coupled approach with the IMU, achieving better loop closure Symposium (LARS), 2021 Brazilian Symposium on Robotics (SBR),
detection by fusing Lidar and IMU data to match scans and and 2021 Workshop on Robotics in Education (WRE), 2021: IEEE,
pp. 216-221.
correct drift. However, tightly coupling can introduce errors if [16] M. Kim, M. Zhou, S. Lee, and H. Lee, "Development of an
either sensor has inaccuracies. Even with loop closure, map Autonomous Mobile Robot in the Outdoor Environments with a
artifacts can still occur due to false positives, false negatives, Comparative Survey of LiDAR SLAM," in 2022 22nd International
sensor quality, environmental factors, and handling of Conference on Control, Automation and Systems (ICCAS), 2022:
IEEE, pp. 1990-1995.
dynamic changes. [17] M. Kaess, H. Johannsson, R. Roberts, V. Ila, J. J. Leonard, and F.
Dellaert, "iSAM2: Incremental smoothing and mapping using the
ACKNOWLEDGMENT Bayes tree," The International Journal of Robotics Research, vol. 31,
no. 2, pp. 216-235, 2012.
The authors thank the MIMOS Berhad for providing [18] G. Wang, X. Gao, T. Zhang, Q. Xu, and W. Zhou, "LiDAR Odometry
facilities and all colleagues at Advance Intelligence Lab for and Mapping Based on Neighborhood Information Constraints for
contributing in this research work. This research is funded by Rugged Terrain," Remote Sensing, vol. 14, no. 20, p. 5229, 2022
Malaysian Ministry of Science, Technology and Innovation