SLAM in Weak Environment Information Applications using Swarm Robots

doi:10.21203/rs.3.rs-4102706/v1

Download PDF

Research Article

SLAM in Weak Environment Information Applications using Swarm Robots

https://doi.org/10.21203/rs.3.rs-4102706/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Simultaneous Localisation and Mapping (SLAM) in weak environment information scenarios where there are only sparse and ambiguous references available presents significant challenges. Feature sparsity and repetitive elements of landmarks impose difficulties in establishing stable tracking and lead perception overlap. The point-line matching based approach presented in this paper provides additional geometric context that makes it easier to distinguish between similar features. This method identifies landmarks as points, extrapolates lines from consecutive point observations, and generates point-line pairs using geometric and appearance-based criteria. The pairs serve as unique identifiers for locations and, hence, effectively solves the identical perceptions problem. To optimize this solution further, we introduce the generation of trajectory graphs, allowing the swarm to refine its map and location estimates iteratively. Experimental results validate that our approach outperforms traditional methods in the challenging weak environment information applications.

Swarm robots

weak environment information

SLAM

Localisation

Point-Line matching

Simultaneous Localisation and Mapping (SLAM) is a critical technology for the autonomous navigation of a swarm of robots. It allows robots to collaboratively build a consistent map of an unknown environment while simultaneously determining their position within it [1]. However, SLAM still poses significant challenges in scenarios with weak environmental information (WEI), where reference features are sparse or ambiguous[2, 3]. The scarcity and repetitiveness of visual landmarks can lead to unstable tracking and inaccurate mapping results, compromising the robustness and reliability of swarm robot SLAM [4].

Existing swarm robot SLAM approaches often face significant challenges in WEI environments due to the perceptual aliasing problem, where different locations may have similar visual appearances[5]. This can lead to false loop closures and inconsistent mapping results due to ambiguity in feature matching. In distributed SLAM algorithms [6], each robot in the swarm constructs its own local map based on limited on-board sensors, which are susceptible to cumulative drift and scale inconsistencies. When fusing these local maps into a global map, perceptual aliasing can lead to incorrect data associations and poor map alignments. On the other hand, consensus-based strategies require the swarm to agree on a common map representation and coordinate frame through communication[7]. However, consensus can be difficult and time consuming in WEI environments where feature correspondences are ambiguous and uncertain.

A number of methods have been proposed to mitigate perceptual aliasing in SLAM for swarm robots. Tsintotas et al [8] introduced a robust method for detecting loop closures based on the spatiotemporal consistency of visual features. They jointly estimated robot poses and identified false loop closures in the swarm using a graph-based optimisation framework. Lajoie et al [9] proposed a swarm robot SLAM system that combines vision and laser range finding data to improve mapping accuracy in challenging environments. The researchers used cascading data association to eliminate unreliable feature matches and maintain consistent global maps across the swarm. Ming Liu et al [10] developed a collaborative SLAM framework for multi robots, exploiting the complementary advantages of different sensors and mapping algorithms. They used an information-theoretic approach to local map fusion and loop closure detection in the presence of perceptual aliasing.

However, even with these efforts, the problem of perceptual aliasing remains a major challenge for swarm robot SLAM in WEI environments. The accuracy and robustness of swarm localisation and navigation can be significantly degraded by incorrect data mappings and inconsistent mapping results[11]. Therefore, developing more effective feature matching and map fusion techniques that can deal with the ambiguity and uncertainty in the visual correspondences between the different viewpoints and robot platforms in the swarm is important.

In this paper, to address the challenges of swarm robot SLAM in a WEI environment, we propose a novel approach based on point-line matching. Our method takes advantage of the complementary nature of point and line features to establish more reliable and discriminative visual correspondences across different points of view. The key idea is the construction of point-line pairs from geometric and appearance constraints, which are used as unique identifiers for scene recognition and loop closure detection. We integrate this matching strategy into a centralised server-based swarm robot system, where individual robots maintain only a compact local map and the server performs global optimisation and map fusion. Our approach effectively solves the perceptual aliasing problem and significantly improves the localisation accuracy and map consistency for swarm robot SLAM with sparse features in WEI environments.

The main contributions of this paper are summarised below:

We propose a visual correspondence method based on point and line matching for the swarm robot SLAM in WEI scenarios. The combination of point and line features provides a richer geometric context for robust data association.

We propose a central server based swarm robot SLAM framework that integrates point and line matching strategies for global map optimisation and robot pose estimation. The centralised architecture solves the inconsistency problem in swarm robot mapping by reducing the amount of on-board computation for individual robots.

We conducted experiments on a public benchmark dataset and a simulation environment to validate the effectiveness and efficiency of our approach. Quantitative and qualitative results show that our method outperforms existing swarm robot SLAM algorithms in WEI scenarios.

The rest of this paper is structured as follows. In Section 2, our swarm robot SLAM method based on point-line matching is described in detail. Section 3 presents experimental results and performance analysis. Finally, Section 4 concludes the paper and discusses future research directions.

In this section, in order to solve the localisation and mapping problem in scenarios with WEI environments, a SLAM method for clustered robots based on point-line matching is presented. First, the architecture and communication protocol of the cluster robot system are presented. Then the point-line matching algorithm for visual feature matching and closed-loop detection is detailed. Finally, the process of global map optimisation and position estimation on the server side is described.

2.1 Server-based centralized swarm robot system

The Swarm Robot SLAM system uses a central server to communicate and compute in an efficient way. The system architecture can be represented by the Fig. 1.

The robot platform is a TurtleBot 3 [12]. Each robot is equipped with a monocular camera and a processor for local feature extraction and position estimation. The robots communicate with a central server via a wireless network to exchange map data and position information. The server maintains a global map of the environment and performs map fusion and optimisation based on the data sent by the robots.

The main advantage of the centralised architecture is that it allows the robot to offload computationally intensive tasks (e.g. global map optimisation and closed-loop detection) to the server [13]. This reduces the on-board processing load on the individual robots, allowing them to focus on localised mapping and navigation. In addition, the centralised server provides a unified coordinate framework for map fusion and robot pose estimation, which helps to mitigate inconsistency and drift problems in distributed SLAM approaches [14].

For achieving the desired algorithmic functionality, we have adopted a communication architecture anchored in the Robot Operating System (ROS)[15]. Within the ROS framework, both a node manager and a parameter manager are established to facilitate efficient data flow between the robots and the centralized server. Subsequent to setting up this communication framework, the mode of information transmission must be selected. Although both wired and wireless options are available, the mobile nature of the robots makes a wireless setup the more pragmatic choice.

In this work, we utilize ROS in conjunction with Wi-Fi to establish a global communication network across the swarm robot system.

2.2 Point-line matching algorithm

The proposed point-line matching algorithm consists of three main stages: feature extraction using oriented fast and rotated brief (ORB)[16] and large-scale direct (LSD)[17] algorithms, generation of point-line pairs based on geometric and appearance-based criteria, and localization and mapping in weak reference information environments using the generated point-line pairs. The algorithm flow can be shown in the following code:

Algorithm 1 Point-line matching for swarm robot SLAM

Input: Robot images

Output: Global map and robot trajectories

1: for each robot in the swarm do

2: Extract point features $\left\{{p}_{i}\right\}$ using ORB

3: Extract line features $\left\{{l}_{j}\right\}$ using LSD

4: Construct point-line pairs $\left\{{p}_{i}, {l}_{j}\right\}$

5: Send point-line pairs to the server

6: end for

7: On the server:

8: Maintain a global database of point-line pairs

9: for each received point-line pair $\left\{{p}_{i}, {l}_{j}\right\}$ do

10: Search for similar pairs in the database using DBoW

11: Perform geometric verification using RANSAC

12: if enough inliers are found then

13: Add a loop closure constraint to the pose graph

14: end if

15: end for

16: Optimize the pose graph to obtain global map and robot trajectories

17: Send updated poses back to the robots

2.3 Generating point-line pairs

The proposed point-line matching method generates point-line pairs to overcome perceptual overlap issues in weak reference information environments. By establishing correspondences between point features and line features, the method improves the accuracy and robustness of the relative pose estimation process.

ORB algorithms are utilized to identify points as landmarks within the environment, serving as discrete locations for robotic navigation. The ORB algorithm extracts keypoints ${p}_{i}$ in an image $I$, where $i$ ranges from 1 to $N,N$ being the total number of keypoints. Each keypoint ${p}_{i}$ is described by its coordinates $\left({x}_{i},{y}_{i}\right)$. Mathematically, this can be expressed as:

$${\text{p}}_{\text{i}}={\text{f}}_{\text{O}\text{R}\text{B}}\left(\text{I}\right), \text{i}=\text{1,2},\dots ,\text{N}$$

LSD techniques extrapolate lines from consecutive point observations, thereby providing a richer representation of the environment. Line segments ${l}_{j}$ are identified within the same visual frame, where $j$ ranges from 1 to $M$, $M$ being the total number of line segments. Each line segment ${l}_{j}$ is defined by its endpoints $\left({x}_{j1},{y}_{j1}\right)$ and $\left({x}_{j2},{y}_{j2}\right)$. Mathematically, this can be formalized as:

$${\text{l}}_{\text{j}}={\text{f}}_{\text{L}\text{S}\text{D}}\left(\text{I}\right), \text{j}=\text{1,2},\dots ,\text{M}$$

Subsequently, a Distributed Bag-of-Words (DBoW)[18] method is employed to establish correspondences between these point and line features across consecutive frames, thereby creating a richer dataset. After feature extraction, DBoW is employed to establish correspondences between these features. The descriptors of keypoints ${p}_{i}$ and line segments ${l}_{j}$ are encapsulated as $d\left({p}_{i}\right)$ and $d\left({l}_{j}\right)$ respectively. Similarity $S$ between descriptors from two consecutive frames ${I}_{t}$ and ${I}_{t-1}$ is computed using Euclidean distance:

$$\begin{array}{rr}\text{S}\left(\text{d}\left({\text{p}}_{\text{i}}^{\text{t}}\right),\text{d}\left({\text{p}}_{\text{i}}^{\text{t}-1}\right)\right)& ={∥\text{d}\left({\text{p}}_{\text{i}}^{\text{t}}\right)-\text{d}\left({\text{p}}_{\text{i}}^{\text{t}-1}\right)∥}_{2}\\ \text{S}\left(\text{d}\left({\text{l}}_{\text{j}}^{\text{t}}\right),\text{d}\left({\text{l}}_{\text{j}}^{\text{t}-1}\right)\right)& ={∥\text{d}\left({\text{l}}_{\text{j}}^{\text{t}}\right)-\text{d}\left({\text{l}}_{\text{j}}^{\text{t}-1}\right)∥}_{2}\end{array}$$

Finally, to enhance the overall reliability of these correspondences, the Random Sample Consensus (RANSAC) algorithm is applied. The algorithm selects a subset of $n$ feature pairs to estimate a model $M$, and based on this model, it classifies inliers $I$ and outliers $O$. Mathematically, this is given by:

$$\begin{array}{c}\text{M}=\text{R}\text{A}\text{N}\text{S}\text{A}\text{C}\left(\text{n}\right)\\ \text{I},\text{O}=\text{c}\text{l}\text{a}\text{s}\text{s}\text{i}\text{f}\text{y}(\text{M},\text{S})\end{array}$$

By considering points as landmarks and lines as extrapolations from consecutive point observations, the environment is represented in a more comprehensive manner. This ensures a feature correspondence that are foundational for subsequent stages of relative pose estimation.

In swarm robot, the algorithm is applied to construct dictionaries of point and line features. The point features encapsulate location and scale information, while line features encompass the endpoint locations. For point features, these are defined by their location and scale, creating a circle with '$p$' as the center and '$r$' as the radius. Line features, in contrast, are determined by the positions of their endpoints. A distinctive word pair is created when a line segment of a line feature intersects with the circle of a point feature. As demonstrated in Fig. 2, point and line 1 can be grouped to create a word pair. However, the combination of point and line 2 does not meet the criteria for forming a word pair.

These dictionaries serve as a rich source of information for the robots, enabling them to better understand their surroundings and coordinate their activities more effectively.

The generation of word pairs from the intersection of point feature circles and line feature segments offers a unique way to interpret and extract information from sparse data. For instance, in a formation where a robot corresponds to a 'point' and its movement trajectory represents a 'line', the intersection could symbolize an event or interaction, such as obstacle avoidance or task collaboration. This study is based on a collaborative vision SLAM algorithm for swarm robots in which, each robot calculates the pose coordinates based on the initial positioning point as the coordinate origin. Therefore, each robot is in a different coordinate system, the pose fusion cannot be performed directly, but the pose transformation matrix for each robot movement has to be calculated based on the pose information of the key frames.

In this system, the key frame positional information of the robot is as follows.

𝐼= timestamp ,𝑡𝑥,𝑡𝑦,𝑡𝑧,𝑞0,𝑞1,𝑞2,𝑞3 (5)

where, the $timestamp$is the timestamp of the keyframe pose, and ${t}_{x},{t}_{y},{t}_{z}$ are the coordinates describing the position of the keyframe in 3D space; ${q}_{0},{q}_{1},{q}_{2},{q}_{3}$ are the unit quaternions describing the rotation of the keyframe in 3D space.

The rotation matrix $R$ of the keyframe can be calculated from the quaternions, and the transformation relations are as follows:

$$\varvec{R}=\left[\begin{array}{lll}1-2{q}_{2}^{2}-2{q}_{3}^{2}& 2{q}_{1}{q}_{2}+2{q}_{0}{q}_{3}& 2{q}_{1}{q}_{3}-2{q}_{0}{q}_{2}\\ 2{q}_{1}{q}_{2}-2{q}_{0}{q}_{3}& 1-2{q}_{1}^{2}-2{q}_{3}^{2}& 2{q}_{2}{q}_{3}+2{q}_{0}{q}_{1}\\ 2{q}_{1}{q}_{3}+2{q}_{0}{q}_{2}& 2{q}_{2}{q}_{3}-2{q}_{0}{q}_{1}& 1-2{q}_{1}^{2}-2{q}_{2}^{2}\end{array}\right]$$

The keyframe panning matrix is as follows:

$$\varvec{t}={\left[{t}_{x},{t}_{y},{t}_{z}\right]}^{\text{T}}$$

As a result, the bitwise transformation matrix of keyframe $i$ is shown below:

$${\varvec{T}}_{i}=\left[\begin{array}{cc}\varvec{R}& \varvec{t}\\ {0}_{1\times 3}& 1\end{array}\right]$$

The robot ${R}_{A}$ has the pose ${\varvec{T}}_{A,i}$ when passing through scene $M$ at one moment ${t}_{i}$, and the robot ${R}_{B}$ has the pose ${\varvec{T}}_{B,j}$ when passing through the same scene at another moment t${t}_{j}$, which is recognized by the system, and the robot ${R}_{A}$ arrives at scene $N$ with the pose ${\varvec{T}}_{A,j}$.

The cumulative pose transformation of robot ${R}_{A}$ from moment ${R}_{A}$ to moment ${t}_{j}$ is:

$${\varvec{T}}_{A,i\to j}={\varvec{T}}_{A,i+1}{\varvec{T}}_{A,i+2}\cdots {\varvec{T}}_{A,j-1}{\varvec{T}}_{A,j}$$

Thus, robot ${R}_{A}$ is transformed with respect to robot ${R}_{B}$ in local coordinates at moment ${t}_{i}$:

$${\varvec{T}}_{A,i}={\left({\varvec{T}}_{A,i\to j}\right)}^{-1}{\varvec{T}}_{A,j}$$

2.4 Localisation and mapping in WEI environments

Algorithm 2: Localization and Mapping

Input: Robot images

Output: Global map and robot trajectories

1: Initialize an empty global map G

2: Initialize an empty pose graph PG

For each robot R in the swarm do:

3: Extract point features $\left\{{p}_{i}\right\}$ using ORB from R's images

4: Extract line features $\left\{{l}_{j}\right\}$ using LSD from R's images

5: Construct point-line pairs $\left\{{p}_{i}, {l}_{j}\right\}$

6: Send $\left\{{p}_{i}, {l}_{j}\right\}$ to the central server

End for

On the central server:

7: Maintain a global database DB of point-line pairs

For each received point-line pair $\left\{{p}_{i}, {l}_{j}\right\}$ from any robot R do:

8: Search DB for similar point-line pairs using DBoW

9: Perform geometric verification for each candidate match using RANSAC

10: If enough inliers are found then

11: Add or update the corresponding node and edge in PG based on $\left\{{p}_{i}, {l}_{j}\right\}$

End if

End for

12: Optimize PG to refine robot poses and construct the global map G

13: Send updated poses and partial global map updates back to the robots

For each robot R in the swarm do:

14: Update R's local map and pose based on the received data from the server

15: If R detects a potential loop closure locally then

16: Send loop closure information to the server

End if

End for

On the central server:

17: Update PG with new loop closure constraints received from robots

18: Re-optimize PG to refine the global map G and robot trajectories

19: Send final updated poses and global map G back to the robots

Output the final global map G and robot trajectories

There are several key steps in the proposed point-line matching algorithm for the swarm robot SLAM. Firstly, the ORB algorithm will be used to extract the point feature P and the LSD algorithm will be used to extract the line feature L. The extracted features will then be matched with a predefined dictionary of point and line features using the DBoW method. Based on the matched features, point-line pairs PL are generated. This provides a robust and informative representation of the scene geometry.

The global map G is then updated based on these estimated poses, integrating the local maps and observations of each robot into a consistent global representation. To improve the accuracy and consistency of the global map, a closed-loop detection is also performed. This detects and corrects accumulated drifts and errors.

The collaborative mapping is achieved through the exchange of data between the robots of the swarm and the central server. Each robot sends its local map, observations and estimated pose to the server, which aggregates and optimises globally. The individual robots are then sent back the updated global map and optimised robot poses.

The algorithm will iterate through these steps until the mapping process is complete, continuously refining and improving the global map and the robot poses as more and more information is captured.

In order to comprehensively evaluate the performance of the proposed point-and-line matching-based SLAM method for clustered robots in WEI scenarios, we conducted a series of experiments in both simulation and real-world environments. The experimental platform is a Turtlebot 3 Waffle Pi robot equipped with an Intel RealSense D435i depth camera (resolution 848x480, frame rate 30fps) and a Raspberry Pi 3B. The robot communicates via Wi-Fi with a central server configured with an Intel i7-8700 processor and 16GB of RAM.

The research presented evaluates the efficiency of swarm robot algorithms weak environment information scenarios using two distinct methods. Firstly, it leverages a premier, publicly available benchmark dataset. Secondly, it conducts real-time tests in a simulated environment. It's noteworthy that most visual SLAM methodologies employ the public benchmark for algorithm assessment.

We selected two representative Gazebo simulation scenarios to evaluate the performance of the swarm robot system in the ROS, a long corridor and a wide open space.

In the point and line matching algorithm, the ORB feature point extraction threshold is set to 1.2, the LSD line segment detection threshold is set to 0.8, the branching factor and DBoW dictionary depth is set to 6 and 10 respectively, the RANSAC interior point threshold is set to 10 and the maximum number of iterations is set to 1000.

The dataset comprises a collection of 50 video sequences captured using a handheld monocular camera. These sequences were filmed in a variety of challenging and multifaceted real-world scenarios, ranging from confined indoor hallways to wide-open outdoor spaces. The videos are recorded at a resolution of 640x360 pixels and maintain a frame rate of 1.5 frames per second.

In this study, The algorithm is rigorously compared with the existing GMapping [19] and other SLAM algorithms. To minimize the impact of random errors, each sequence is independently run 10 times, with the TUM dataset notably being executed 500 times[20].

The study focuses on three main areas of comparison: trajectory error, tracking time, and local map quality.

3.1 Based on experimental results and analysis of open benchmark

In the current study, three specific sequences—S01, S02, and S03—from the TUM dataset were meticulously chosen for experimentation. To mitigate the impact of stochastic variations, each experimental set was conducted through a series of ten iterations, and the results were subsequently averaged.

Table 1

The Average Trajectory Error Co-Location Based On TUM Dataset
Robot	Our method	GMapping
Robot A (TUM S01)	0.212	0.543
Robot B (TUM S02)	0.255	0.492
Robot C (TUN S03)	0.246	0.472

Based on the data in the Table 1, Our mathod performs better than GMapping in all tested cases. For example, Robot A has an average trajectory error of 0.212 with Our mathod, which is much lower than the 0.543 error with GMapping. Similar improvements are seen for Robots B and C as well. In terms of co-localisation, Our mathod also outperforms with an error of 0.183 compared to GMapping's 0.467. Overall, the data suggests that Our mathod is more accurate and reliable.

Table 2

The Average Tracking Time (in seconds) Based on TUM Dataset
Robot	Our mathod	GMapping
Robot A (TUM S01)	0.75	1.20
Robot B (TUM S02)	0.85	1.12
Robot C (TUN S03)	0.70	1.10

Based on the average tracking time data from the TUM dataset for three different robots in Table 2, Our mathod consistently outperforms GMapping, taking less time to complete tracking across all tested scenarios. The efficiency gains range from 0.35 to 0.45 seconds per robot, indicating a significant improvement in real-time performance for our mathod.

3.2 Swarm robots tracks and map fusion effect

In an simulation environment designed to evaluate the performance of our mathod and GMapping algorithms, a swarm of robots embarks on randomly generated routes. These robots are programmed with the aforementioned SLAM algorithms to navigate and map the space. Critical metrics such as trajectory, tracking time, and map data are meticulously recorded for each robot, and any discrepancies in localisation are quantified.

The trajectory maps of the three robots are shown in Fig. 3, where Robot A and Robot B are shown on the left and Robot C is shown on the right.

The algorithm shows a co-localisation accuracy performance improvement of 60.8%, with a trajectory error of 0.183 m compared to 0.467 m in GMapping. Figure 3(a) displays a fused global map that includes trajectories from Robot A and Robot B. Robot A's trajectory is represented by a dashed line and Robot by a solid line. Figure 3(b) presents the trajectory map for Robot C, who does not share the same space with Robots A and B at the same time. The location recognition module incorporates co-visual constraints both within individual trajectories and between different robots' trajectories, which aids in co-localisation and map building.

3.3 Summary

Based on the experimental results, the proposed our mathod algorithm reduces the average trajectory error to 0.183 m from 0.467 m in GMapping, leading to an increase in co-localisation accuracy by 60.8%. The tracking time also decreases consistently across all test cases, showing better real-time performance.

The algorithm addresses the issue of identical perceptions among swarm robots. By incorporating co-visual constraints within and between robots' trajectories, the system differentiates identical perceptions and enhances co-localisation and map building. The results confirm that our mathod effectively tackles the challenges of trajectory error, tracking time, and identical perceptions in swarm robotics.

In our study, we present an innovative approach to overcome the challenges faced by swarming robotic systems when performing SLAM in environments with minimal and ambiguous cues. Our approach builds a dictionary of point and line features using a centralised framework and a DBoW model. This point-line pairing strategy effectively reduces the problem of perceptual ambiguity, thereby improving the accuracy and stability of robot swarm localisation and mapping. Our empirical results show that our technique is highly effective compared to traditional methods. This highlights its applicability in challenging environments where SLAM of robot swarms is critical.

In conclusion, our novel point-line matching-based SLAM method for swarm robots systems has shown promising results in weak environments with sparse and ambiguous reference information. Further research into relative posture calculations and the integration of machine learning techniques holds significant potential for advancing the state of the art in swarm robots SLAM and unlocking new applications in WEI environments.

The work in this paper is sponsored by the EU H2020 project “5G Enhance Robot Autonomy” under grant agreement number 101016681.

Author Contribution

All authors reviewed the manuscript.

Bresson G, Alsayed Z, Yu L, Glaser S (2017) Simultaneous Localization and Mapping: A Survey of Current Trends in Autonomous Driving. IEEE Transactions on Intelligent Vehicles, Intelligent Vehicles, IEEE Transactions on, IEEE Trans. Intell. Veh 2 (3): 194–220. 10.1109/TIV.2017.2749181
Muhammad H, Ishikawa Y, Sekiguchi K (2024) and Kenichiro Nonaka. Online Simultaneous Localization and Mapping with Parallelization for Dynamic Line Segments Based on Moving Horizon Estimation. Artificial Life and Robotics, March, 1–15. 10.1007/s10015-024-00937-8
Mur-Artal R, Montiel JMM, Tardos JD (2015) ORB-SLAM: A Versatile and Accurate Monocular SLAM System. IEEE Transactions on Robotics, Robotics, IEEE Transactions on, IEEE Trans. Robot 31 (5): 1147–63. 10.1109/TRO.2015.2463671
Lajoie P, Ramtoula B, Chang Y, Carlone L, Beltrame G (2020) DOOR-SLAM: Distributed, Online, and Outlier Resilient SLAM for Robotic Teams. IEEE Robotics and Automation Letters, Robotics and Automation Letters, IEEE, IEEE Robot. Autom. Lett 5 (2): 1656–63. 10.1109/LRA.2020.2967681
Angeli A, Filliat D, Doncieux S (2008) and J-A. Meyer. Fast and Incremental Method for Loop-Closure Detection Using Bags of Visual Words. IEEE Transactions on Robotics, Robotics, IEEE Transactions on, IEEE Trans. Robot 24 (5): 1027–37. 10.1109/TRO.2008.2004514
Tian Y, Chang Y, Herrera Arias F, Nieto-Granda C, How J, Carlone L (2022) Kimera-Multi: Robust, Distributed, Dense Metric-Semantic SLAM for Multi-Robot Systems. IEEE Transactions on Robotics, Robotics, IEEE Transactions on, IEEE Trans. Robot 38 (4): 2022–38. 10.1109/TRO.2021.3137751
Endres F, Hess J, Engelhard N, Sturm J, Cremers D (2012) and Wolfram Burgard. An Evaluation of the RGB-D SLAM System. 2012 IEEE International Conference on Robotics and Automation, Robotics and Automation (ICRA), 2012 IEEE International Conference On, May, 1691–96. 10.1109/ICRA.2012.6225199
Tsintotas KA, Bampis L, Gasteratos A (2022) The Revisiting Problem in Simultaneous Localization and Mapping: A Survey on Visual Loop Closure Detection. IEEE Transactions on Intelligent Transportation Systems, Intelligent Transportation Systems, IEEE Transactions on, IEEE Trans. Intell. Transport. Syst 23 (11): 19929–53. 10.1109/TITS.2022.3175656
Lajoie P-Y, Beltrame G (2024) Swarm-SLAM: Sparse Decentralized Collaborative Simultaneous Localization and Mapping Framework for Multi-Robot Systems. IEEE Rob Autom Lett 9(1):475–482. 10.1109/LRA.2023.3333742
Ming Liu R, Siegwart, Deutsch I (2016) A Framework for Multi-Robot Pose Graph SLAM. 2016 IEEE International Conference on Real-Time Computing and Robotics (RCAR), June. https://research.ebsco.com/linkprocessor/plink?id=dae61994-e632-3043-a9ac-121c2b72bb62
Kaveti P, Gupta A, Giaya D, Karp M, Keil C, Nir J, Zhang Z (2023) and Hanumant Singh. Challenges of Indoor SLAM: A Multi-Modal Multi-Floor Dataset for SLAM Evaluation. 2023 IEEE 19th International Conference on Automation Science and Engineering (CASE), Automation Science and Engineering (CASE), 2023 IEEE 19th International Conference On, August, 1–8. 10.1109/CASE56687.2023.10260618
Robotis-Git. Robotis-Git/Emanual: Welcome to the Robotis e-Manual ! The e-Manual Page Rendered from This Repository Is Available for Everyone. Just Simply Click the Provided Link Below:). GitHub. Accessed March 14, (2024) https://github.com/ROBOTIS-GIT/emanual
de Koning C, and Anahita Jamshidnejad (2023) Hierarchical Integration of Model Predictive and Fuzzy Logic Control for Combined Coverage and Target-Oriented Search-and-Rescue via Robots with Imperfect Sensors. J Intell Robotic Systems: Special Sect Unmanned Syst 107(3). 10.1007/s10846-023-01833-2
Dah-Achinanon U, Bajestani SE, Marjani P-Y, Lajoie, and Giovanni Beltrame (2023) Search and Rescue with Sparsely Connected Swarms. Auton Robots 47(7):849–863. 10.1007/s10514-022-10080-7
Quigley M, Conley K, Gerkey B, Faust J, Foote T, Leibs J et al (2009) Ros: an open-source robot operating system, ICRA workshop on open source software, vol. 3, no. 3.2, pp. 5, Japan
Campos C, Elvira R, Juan J, Gomez Rodriguez JMM, Montiel, Tardos JD (2021) ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual–Inertial, and Multimap SLAM. IEEE Trans Robot 37(6):1874–1890. 10.1109/TRO.2021.3075644
Engel J, Stuckler J (2015) and Daniel Cremers. Large-Scale Direct SLAM with Stereo Cameras. 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Intelligent Robots and Systems (IROS), 2015 IEEE/RSJ International Conference On, September, 1935–42. 10.1109/IROS.2015.7353631
Le Q (2014) and Tomas Mikolov. Distributed representations of sentences and documents. International conference on machine learning. PMLR
Grisetti G, Stachniss C, Burgard W (2007) Improved Techniques for Grid Mapping With Rao-Blackwellized Particle Filters. IEEE Transactions on Robotics, Robotics, IEEE Transactions on, IEEE Trans. Robot 23 (1): 34–46. 10.1109/TRO.2006.889486
Walch F, Hazirbas C, Leal-Taixe L, Sattler T, Hilsenbeck S, Cremers D (2017) Image-Based Localization Using LSTMs for Structured Feature Correlation. 2017 IEEE International Conference on Computer Vision (ICCV), Computer Vision (ICCV), 2017 IEEE International Conference on, ICCV, October, 627–37. 10.1109/ICCV.2017.75

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

SLAM in Weak Environment Information Applications using Swarm Robots

Status:

Version 1

Abstract

Figures

1 Introduction

2 Based on relative point-line matching for scene recognition

2.1 Server-based centralized swarm robot system

2.2 Point-line matching algorithm

2.3 Generating point-line pairs

2.4 Localisation and mapping in WEI environments

3 The experimental content design

3.1 Based on experimental results and analysis of open benchmark

3.2 Swarm robots tracks and map fusion effect

3.3 Summary

4 Conclusions

Declarations

Author Contribution

References

Additional Declarations

Status:

Version 1