PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Simultaneous Localization and Mapping with an Envision in 6G Wireless Networks

Baichuan Huang, Jun Zhao, Jingbin Liu

arXiv:1909.05214v4cs.RO

TL;DR

SLAM research still faces the challenge of balancing accuracy, robustness, and limited platform resources. This survey reviews Lidar and visual SLAM, their fusion, and future directions including 6G, and concludes that Lidar–vision fusion offers a balance of reliability and versatility.

  • Problem

    A key open challenge is balancing accuracy, robustness, and limited platform resources in SLAM.

    Source: C. Challenge and Future [297] · p. 9

  • Method

    The survey reviews Lidar and visual SLAM sensors and systems, visual-inertial odometry, their fusion, and future directions including 6G.

    Source: §VI · p. 10

  • Results

    The paper describes Lidar–vision fusion as a balanced way to achieve reliability and versatility.

    Source: §VI · p. 10

  • Takeaways & Limitations

    The overview provides a user guide for new researchers in SLAM.

    Source: §I · p. 1

Abstract

from arXiv · show

Simultaneous Localization and Mapping (SLAM) achieves the purpose of simultaneous positioning and map construction based on self-perception. The paper makes an overview in SLAM including Lidar SLAM, visual SLAM, and their fusion. For Lidar or visual SLAM, the survey illustrates the basic type and product of sensors, open source system in sort and history, deep learning embedded, the challenge and future. Additionally, visual inertial odometry is supplemented. For Lidar and visual fused SLAM, the paper highlights the multi-sensors calibration, the fusion in hardware, data, task layer. The open question and forward thinking with an envision in 6G wireless networks end the paper. The contributions of this paper can be summarized as follows: the paper provides a high quality and full-scale overview in SLAM. It's very friendly for new researchers to hold the development of SLAM and learn it very obviously. Also, the paper can be considered as a dictionary for experienced researchers to search and find new interesting orientation.

I. INTRODUCTION · II. LIDAR SLAM · A. Lidar Sensors

SLAM jointly estimates robot location and constructs an environment map, an interdependent problem that has evolved from filtering toward graph-based, multisensor systems. This survey introduces SLAM’s development and scope, including Lidar SLAM and its sensor types and products.

I. INTRODUCTION

  • SLAM couples localization and mapping: precise motion estimates require an accurate map, while accurate mapping requires precise knowledge of the robot’s sensing locations.Source: §I · p. 1
  • SLAM’s localization challenge is the interdependence of knowing the robot’s pose and building its map; applied EKF to incrementally estimate robot pose and landmark positions.GPS works well only outdoors, IMUs accumulate error, and wireless localization faces a cost–accuracy tradeoff; the passage describes SLAM with Lidar, cameras, IMUs, and other sensors as emerging.Source: §I · p. 1
  • The survey describes SLAM’s shift from KF, EKF, and PF filtering to dominant graph-based optimization, alongside a move from single-thread to multi-thread systems.Source: §I · p. 1
  • The survey covers Lidar, visual, and fused SLAM, including sensors, open-source systems, deep learning, challenges, visual-inertial odometry, and a 6G outlook.It presents itself as a user guide for new SLAM researchers.Source: §I · p. 1

II. LIDAR SLAM

  • After sonar-equipped SLAM using multiple servo-mounted sensors and EKF, the advent of Lidar made SLAM systems more reliable and robust.Source: §II · p. 1

A. Lidar Sensors

  • Lidar sensors are categorized by beam count as 2D or 3D, and by production process as mechanical, hybrid solid-state, or solid-state using phased-array or flash technology.Source: A. Lidar Sensors · p. 1
  • Examples include Velodyne mechanical and hybrid models, Ouster mechanical units with 16–128 channels, and Ibeo mechanical models and its Valeo-co-developed Scala hybrid unit.Source: A. Lidar Sensors · pp. 1–2
  • The sensor overview also lists SLAMTEC’s low-cost RPLIDAR products and Quanergy’s solid-state S3 and mechanical M8; it forecasts a market trend toward miniaturized, lightweight solid-state Lidar.Source: A. Lidar Sensors · p. 2

B. Lidar SLAM System · 1) 2D SLAM • · 2) 3D SLAM

The section surveys LiDAR SLAM systems across 2D and 3D, describing their methods, inputs, and capabilities.

B. Lidar SLAM System

1) 2D SLAM •

  • Gmapping is a widely used robot SLAM package combining an RBPF method with scan matching, and it builds a grid map as an improvement on FastSLAM [8][6][9][10].Source: 1) 2D SLAM • · p. 2
  • HectorSlam combines 2D SLAM, 3D navigation, scan matching, and inertial sensing [11].Source: 1) 2D SLAM • · p. 2
  • KartoSLAM is graph-based [12], LagoSLAM minimizes a nonlinear non-convex cost function [13], and CoreSLAm aims for minimal performance loss [14].Source: 1) 2D SLAM • · p. 2
  • Cartographer uses sub-maps and loop closure for product-grade performance and supports 2D and 3D SLAM across platforms and sensor configurations [15].Source: 1) 2D SLAM • · p. 2

2) 3D SLAM

  • LOAM provides real-time state estimation and mapping with 3D LiDAR, with variants for back-and-forth spinning and continuous-scanning 2D LiDAR [16].Source: 2) 3D SLAM · p. 2
  • Lego-Loam uses point clouds from a horizontally placed Velodyne VLP-16 and optional IMU data to produce real-time 6D pose estimates with global optimization and loop closure [17].Source: 2) 3D SLAM · p. 2
  • IMLS-SLAM uses only 3D LiDAR in a scan-to-model matching framework to provide low-drift SLAM [18].Source: 2) 3D SLAM · p. 2

3) Deep Learning With Lidar SLAM … 3) Adversarial Sensor Attack

The section surveys deep-learning methods for LiDAR feature extraction, detection, recognition, segmentation, and localization, then identifies cost, changing environments, and adversarial attacks as challenges for LiDAR SLAM.

3) Deep Learning With Lidar SLAM

  • Deep-learning LiDAR work spans global-descriptor place recognition and end-to-end 3D detection, including PointNetVLAD [19], VoxelNet [20], and sensor-fusion framework MV3D.Source: 3) Deep Learning With Lidar SLAM · p. 2
  • Methods include point-cloud classification and segmentation (PointNet [30]), hierarchical context learning (PointNet++ [31]), temporally consistent LiDAR segmentation (DeepTemporalSeg [43]), and segment-based mapping and loop closure (SegMap [33]; SegMatch).Source: 3) Deep Learning With Lidar SLAM · p. 2
  • L3-Net provides centimeter-level LiDAR localization, while SuMa++ [47] uses semantic labels to build enriched surfel maps and constrain projective scan matching.Source: 3) Deep Learning With Lidar SLAM · p. 3

1) Cost and Adaptability

  • LiDAR offers 3D information, robustness to night and light changes, and up to 360-degree viewing, but high development costs motivate miniaturization, solid-state designs, and greater reliability and adaptability.Source: 1) Cost and Adaptability · p. 3

2) Low-Texture and Dynamic Environment

  • Low-texture and dynamic environments challenge LiDAR SLAM; uses IMU assistance for 2D SLAM, while incorporates time into mapping, and robust map updating remains an open concern.Source: 2) Low-Texture and Dynamic Environment · p. 3

3) Adversarial Sensor Attack

  • LiDAR adversarial attacks can spoof distance estimates or incapacitate sensing in a direction; optimized spoofing attacks [51] achieved around 75% success, motivating defenses for LiDAR SLAM.Source: 3) Adversarial Sensor Attack · p. 3

III. VISUAL SLAM · A. Visual Sensors

The section outlines visual SLAM’s processing pipeline and surveys camera types, their capabilities, and limitations.

III. VISUAL SLAM

  • More powerful CPU/GPU processing and cheaper, lighter, versatile cameras accompany visual SLAM’s rapid development; systems can run on embedded devices and smartphones [52][53][54][55][56].The survey contrasts visual systems’ smaller, cheaper form with lidar systems.Source: §III · p. 3
  • Visual SLAM combines camera/IMU sensing, visual or visual-inertial odometry, backend optimization and loop closure, mapping, and relocalization; Fourier–Mellin VO appears in feature-poor ground-facing-camera examples [60][61].Relocalization is described as an additional module for stable and accurate visual SLAM.Source: §III · p. 3

A. Visual Sensors

  • Visual SLAM commonly uses monocular, stereo, RGB-D, and event cameras.The survey identifies cameras as the most-used visual SLAM sensors.Source: A. Visual Sensors · p. 3
  • Monocular cameras cannot recover real depth, creating scale ambiguity; monocular SLAM also requires initialization and faces drift.Source: A. Visual Sensors · p. 3
  • Stereo cameras have a known baseline and can recover depth through calibration, correction, matching, and calculation, but this process consumes substantial resources.Source: A. Visual Sensors · p. 3
  • RGB-D cameras output per-pixel depth, using stereo, structured light, or time-of-flight (TOF) technology.Structured light uses infrared patterns and an IR camera; TOF measures laser flight time to calculate distance.Source: A. Visual Sensors · p. 3
  • Event cameras asynchronously measure per-pixel brightness changes, offering 140 dB versus 60 dB dynamic range, microsecond-order temporal resolution, low power use, and no motion blur [63].The survey says they can perform better than traditional cameras at high speed and high dynamic range.Source: A. Visual Sensors · p. 3

B. Visual SLAM System · 1) Sparse Visual SLAM •

Visual SLAM methods are divided into direct approaches, which produce semi-dense or dense maps, and feature-based approaches, which produce sparse maps. The sparse-system overview spans monocular, stereo, and broader camera support, alongside systems with distinct tracking, optimization, and relocalization capabilities.

B. Visual SLAM System

1) Sparse Visual SLAM •

  • MonoSLAM is the first real-time monocular SLAM system and is based on an EKF [71].Source: 1) Sparse Visual SLAM • · p. 4
  • PTAM was the first system to parallelize tracking and mapping, and introduced Bundle Adjustment and key frames [54].A later version added a simple, effective relocalization method.Source: 1) Sparse Visual SLAM • · p. 4
  • ORB-SLAM uses Tracking, Local Mapping, and Loop Closing threads; ORB-SLAM v2 supports monocular, stereo, and RGB-D cameras [52][75].CubemapSLAM adapts ORB-SLAM to monocular fisheye cameras [76], while Visual Inertial ORB-SLAM describes IMU initialization and joint visual optimization [77][78].Source: 1) Sparse Visual SLAM • · p. 4
  • proSLAM is a lightweight stereo visual SLAM system designed to be easy to understand [79].Source: 1) Sparse Visual SLAM • · p. 4
  • ENFT-sfm efficiently matches feature-point correspondences across one or multiple video sequences, and its updated ENFT-SLAM version can run at large scale [80].Source: 1) Sparse Visual SLAM • · p. 4
  • OpenVSLAM is an indirect sparse-feature system that supports perspective, fisheye, equirectangular, and user-designed camera models.Source: 1) Sparse Visual SLAM • · p. 4
  • TagSLAM uses AprilTag fiducial markers and provides a front end to the GT-SAM factor-graph optimizer for designing experiments [82].The passage also lists UcoSLAM as similar work [83].Source: 1) Sparse Visual SLAM • · p. 4

2) SemiDense Visual SLAM • · 3) Dense Visual SLAM •

The survey reviews semi-dense visual odometry methods and dense visual SLAM systems, describing their sensor types, mapping capabilities, and distinguishing techniques. The methods range from direct or sparse image alignment to dense reconstruction and mapping of deformable scenes.

2) SemiDense Visual SLAM •

  • LSD-SLAM uses direct monocular tracking on Lie algebra, with stereo and omnidirectional extensions [86]; DSO [91] uses direct sparse odometry without feature detection or description.Other omnidirectional-camera work is cited as.Source: 2) SemiDense Visual SLAM • · p. 4
  • SVO uses sparse model-based alignment for speed and extends to multiple camera types [78]; CNN-SVO adds single-image depth prediction [89], while event-based EVO resists motion blur and handles high-dynamic-range illumination changes.Other event-camera semi-dense SLAM and visual odometry systems are cited as and [95].Source: 2) SemiDense Visual SLAM • · p. 4

3) Dense Visual SLAM •

  • DTAM reconstructs 3D in real time through a globally spatially regularized energy functional [97], while MLM SLAM estimates dense depth and smooths it online without a GPU.DTAM uses a novel non-convex optimization framework; MLM SLAM uses multi-resolution depth estimation.Source: 3) Dense Visual SLAM • · p. 4
  • Kinect Fusion was among the first depth-camera 3D reconstruction systems [100]; DVO uses entropy-based keyframe selection and g2o-based loop closure [102][103], while RGBD-SLAM-V2 reconstructs accurate dense models without other sensors.InfiniTAM provides real-time 3D reconstruction using a CPU on Linux, iOS, and Android.Source: 3) Dense Visual SLAM • · p. 4
  • Kintinuous creates globally consistent point and mesh reconstructions in real time [106][107]; Elastic Fusion captures dense, globally consistent surfel maps of room-scale environments [116], and stereo SOFT-SLAM supports large loop closing [122]..Source: 3) Dense Visual SLAM • · pp. 4–5
  • RTAB-MAP supports visual and Lidar SLAM but is difficult to use as a basis for upper algorithms [108][109][110][111]; KO-Fusion combines dense RGB-D SLAM with wheeled-robot kinematic and odometry measurements..Source: 3) Dense Visual SLAM • · pp. 4–5
  • Dynamic Fusion reconstructs non-rigidly deforming scenes densely in real time based on Kinect Fusion; VolumeDeform also reconstructs non-rigid scenes in real time but is not open source.Related work includes Fusion4D.Source: 3) Dense Visual SLAM • · p. 4
  • Bundle Fusion recovers tracking after gross failures and re-estimates its 3D model in real time to maintain global consistency..Source: 3) Dense Visual SLAM • · p. 5

4) Visual Inertial Odometry SLAM

The section presents VIO as a popular research direction amid limitations in visual SLAM, then surveys VIO systems, products, and learning-based approaches. It describes monocular initialization and scale problems alongside challenges affecting stereo and RGB-D cameras.

Visual Inertial Odometry SLAM

  • Monocular visual SLAM faces initialization, scale ambiguity, and scale drift, while stereo and RGB-D systems still encounter fast motion, limited field of view, computation, occlusion, feature loss, dynamic scenes, and changing light.The passage notes that Global Shuttle or fisheye and panoramic cameras can address fast movement.Source: 4) Visual Inertial Odometry SLAM · p. 5
  • VIO includes products such as Tango, Dyson 360 Eye, and HoloLens, as well as ARKit, ARCore, Insideout, event-camera systems, and multi-IMU mi-VINS.The survey also mentions RKSLAM for fast motion and strong rotation in AR applications.Source: 4) Visual Inertial Odometry SLAM · p. 5
  • Filter-based systems include loosely coupled SSF and tightly coupled MSCKF and ROVIO, with ROVIO tracking 3D landmarks and image patches.MSCKF-VIO is an open-source version of similar work, and ROVIO supports monocular cameras.Source: 4) Visual Inertial Odometry SLAM · p. 5
  • Optimization-based systems include keyframe-based OKVIS, real-time VINS [53][143], parallel local and global bundle adjustment in ICE-BA, and multi-session mapping tools in maplab.VINS-Fusion supports multiple visual-inertial sensor types, online spatial and temporal calibration, and visual loop closure; maplab also supports map merging, batch optimization, and dense reconstruction.Source: 4) Visual Inertial Odometry SLAM · p. 5
  • Learning-based VIO networks can operate without IMU intrinsic parameters or camera–IMU extrinsic calibration, while another network avoids camera–IMU calibration.Source: 4) Visual Inertial Odometry SLAM · p. 5

5) Deep Learning with Visual SLAM

The section surveys how deep learning and semantic information are incorporated into visual SLAM across feature detection, recognition, depth and scale recovery, pose estimation, long-term localization, and dynamic-scene handling.

Feature & Detection

  • Learned keypoints, descriptors, planes, and 3D object models extend visual SLAM from feature tracking toward denser, more semantic mapping, including in low-texture scenes [159][161][165].The surveyed approaches include Pop-up SLAM’s per-image plane measurements, SuperPoint’s self-supervised detector and descriptors, GCNv2 keypoints, and CubeSLAM’s object-level mapping.Source: 5) Deep Learning with Visual SLAM · pp. 5–6

Recognition & Segmentation

  • Recognition and segmentation methods transfer image semantics into 3D maps, using prior object models, RGB-D fusion, and learned semantic or instance segmentation [173][181].Examples include Semanticfusion’s combination of CNNs with ElasticFusion, and 3D-SIS’s instance segmentation for commodity RGB-D scans.Source: 5) Deep Learning with Visual SLAM · p. 6

Recovery Scale

  • Deep models address monocular depth and scale recovery through learned depth, pose, optical flow, and geometric constraints, including jointly unsupervised estimation in GeoNet [196][197].CodeSLAM jointly optimizes a single-image depth map with pose variables; the survey also notes depth-estimation work for event-camera monocular and stereo settings [206].Source: 5) Deep Learning with Visual SLAM · p. 6

Pose Output & Optimization

  • Pose approaches range from single-image 6-DOF prediction with PoseNet to recurrent monocular VO with DeepVO and semantic point tracking with VSO.The section also describes VInet’s motion estimation in visual-inertial odometry and MID-Fusion’s object-oriented pose estimation and model fusion [127].Source: 5) Deep Learning with Visual SLAM · p. 6

Long-term Localization

  • Long-term localization methods use semantic landmarks, learned loop-closure embeddings, and semantic graph matching; X-View supports global localization across drastically different viewpoints [219][221].The surveyed methods also use semantic information for descriptor-based localization and represent object-class and landmark-association hypotheses [222].Source: 5) Deep Learning with Visual SLAM · p. 7

Dynamic SLAM

  • Dynamic-SLAM systems use semantic information, object detection, or static-map reconstruction to improve operation in moving scenes, while SimVODIS jointly outputs depth and relative pose and detects and segments objects [229][234].Examples include MaskFusion’s object-aware semantic RGB-D mapping and StaticFusion’s simultaneous moving-object detection and background reconstruction [230].Source: 5) Deep Learning with Visual SLAM · p. 7

C. Challenge and Future … 4) Software & hardware

The paper frames future SLAM as a challenge of robust operation across difficult conditions and environments, while emphasizing sensor fusion, semantic understanding, and tighter integration of algorithms and hardware.

1) Robustness and Portability

  • Visual SLAM must address illumination changes, dynamic scenes, fast motion, vigorous rotation, and low-texture environments; global-shutter cameras and event cameras are proposed as ways to improve robustness.The passage says event cameras can produce up to one million events per second, supporting very fast motion and high dynamic range.Source: 1) Robustness and Portability · p. 7
  • Potential robustness strategies include semantic, edge, plane, or surface features, reduced feature dependence, and combinations of direct tracking and machine learning.The passage also prefers precise mathematical formulations for SfM/SLAM over navigation functions implicitly learned from data.Source: 1) Robustness and Portability · p. 7
  • Future SLAM may target smartphones and embedded platforms such as UAVs, alongside detailed 3D reconstruction and deep-learning-based scene understanding, while balancing real-time operation and accuracy.Solutions for dynamic, unstructured, complex, uncertain, and large-scale environments remain unexplored.Source: 1) Robustness and Portability · p. 7

2) Multiple Sensors Fusion

  • SLAM systems can combine complementary sensors, such as visual and IMU data in mobile-phone VIO, and sensor choice depends on the environment and required map type.DeLS-3D combines camera video, GPS/IMU motion sensors, and a 3D semantic map for robustness and efficiency.Source: 2) Multiple Sensors Fusion · p. 7

3) Semantics SLAM

  • Semantic SLAM uses deep-learning-based object recognition and segmentation to improve environmental perception and support global optimization, loop closure, and relocalization.The passage contrasts this with geometric-feature approaches, including PL-SLAM and StructSLAM, and describes semantic SLAM as a route toward high-precision real-time positioning at large scale.Source: 3) Semantics SLAM · p. 7

4) Software & hardware

  • Because SLAM is an integrated technology dependent on both software and hardware, future systems are expected to combine algorithms more deeply with sensors.Domain-specific processors and integrated sensor modules, rather than general processors and separate sensors such as a standalone camera, are identified as promising.Source: 4) Software & hardware · p. 7

IV. LIDAR AND VISUAL SLAM SYSTEM … C. Challenge and Future [297]

The survey reviews calibration and fusion across sensor modalities, then identifies open challenges involving multi-sensor coordination, hardware, crowdsourcing, mapping, robustness, and learning. It frames these as questions for future SLAM systems and applications.

A. Multiple Sensors Calibration •

  • Camera–IMU calibration includes multiple-camera, visual-inertial, and rolling-shutter calibration; systems also provide online calibration or estimate extrinsics and time offsets [143][140][242].IMU-TK calibrates IMU internal parameters [244], and proposes an end-to-end monocular VIO network fusing camera and IMU data.Source: A. Multiple Sensors Calibration • · pp. 7–8
  • Other calibration work spans synchronized RGB-depth benchmarks, multi-camera SLAM [249][250], lidar–IMU methods [251][252][253][254], and camera–lidar optimization, pipelines, and learned extrinsic estimation [256][257][259].LIMO extracts depth from lidar measurements for camera feature tracks and motion estimation.Source: A. Multiple Sensors Calibration • · p. 8

B. Lidar and Visual Fusion •

  • At the hardware layer, Pandora integrates 40-beam lidar, five color cameras, and recognition software to ease synchronization; at the data layer, sparse high-precision lidar depth and dense low-precision camera depth motivate upsampling and completion methods [272][273][274][275].LIC-Fusion fuses IMU measurements, sparse visual features, and extracted lidar points [278].Source: B. Lidar and Visual Fusion • · p. 8
  • Task-layer fusion spans perception and moving-object detection, odometry and mapping, loop closure, and tracking, including V-Loam’s lidar refinement of visual odometry and VIL-SLAM’s tightly coupled stereo VIO and lidar mapping [280][285][286][287][288].The survey also describes depth-assisted VO, monocular visual SLAM aided by laser distances, and learning-based camera–lidar detection, recognition, and localization [291][292][26][293][295][296].Source: B. Lidar and Visual Fusion • · p. 8

C. Challenge and Future [297]

  • Future SLAM must unify sensor data types, timestamps, and coordinate systems, while addressing cross-sensor modeling and optimization and improving integrated hardware, robustness, and deployment from front-end sensors to SDKs.The survey notes that malfunction, off-nominal conditions, or aging can make sensor noise and bias depart from the assumed noise model; it also identifies the lack of suitable integrated SLAM chips and hardware.Source: C. Challenge and Future [297] · p. 9
  • Decentralized multi-robot visual SLAM raises crowdsourcing and privacy challenges, for which the survey suggests differential privacy as a possible aid [299][300].Source: C. Challenge and Future [297] · p. 9
  • The survey leaves open which map type best supports robots and how long-term maps should be updated or outdated information discarded.Source: C. Challenge and Future [297] · p. 9
  • No current SLAM system covers all scenarios; future systems must reduce parameter tuning, support cross-condition loop closure, handle failure and risk, and balance accuracy, robustness, and platform resources.The survey favors appearance-based over feature-based methods, with semantic information, to help close loops across day/night sequences and seasons; it also asks whether end-to-end learning will dominate SLAM.Source: C. Challenge and Future [297] · p. 9

V. AN ENVISION IN 6G WIRELESS NETWORKS · A. The advantages of 6G Wireless Networks

The section presents 6G’s high data rates and low latency as a response to SLAM needs that 5G cannot meet, and outlines further network capabilities including ubiquitous connectivity and AI-enabled services.

V. AN ENVISION IN 6G WIRELESS NETWORKS

  • SLAM for robots and autonomous cars needs higher data rates and lower latency than 5G can provide; 6G uses 100 GHz to 3 THz frequencies to offer greater rates than 5G’s 100 Gbps.The passage identifies these communication demands as a motivation for 6G.Source: §V · p. 9
  • The survey describes THz as the band between microwave frequencies below THz and optical frequencies above THz, and says 6G THz is unaffected by light changes and NLOS.It also says 6G will not need support such as 5G’s MIMO represented as mmWave communications.Source: §V · p. 9

A. The advantages of 6G Wireless Networks

  • The cited passage lists low latency as a capability and specifies less than 1 msec end-to-end latency.Source: A. The advantages of 6G Wireless… · p. 9
  • The cited passage lists the data-rate metric as high data rates, with an upper value of 1 Tbps; it does not assign that figure to 6G.Source: A. The advantages of 6G Wireless… · p. 9
  • 6G is envisioned to connect a global network of intelligent things and smart surfaces or environments, from walls and roads to entire buildings.The passage frames this as ubiquitous connection.Source: A. The advantages of 6G Wireless… · p. 9
  • AI and RISs are described as enabling an intelligent network beyond classical big-data analytics and edge computation, with communication, computation, and storage resources.The passage also presents 6G as a platform for these resources and cites.Source: A. The advantages of 6G Wireless… · p. 9

B. 6G in Simultaneous Localization and Mapping · VI. CONCLUSION

The paper surveys Lidar, visual, and fused SLAM and outlines future sensor-based research alongside a 6G radio-based vision. It presents 6G as a networked sensing and localization platform with prospective gains in accuracy, coverage, and remote computation.

B. 6G in Simultaneous Localization and Mapping

  • Radio-based SLAM can use satellite, cellular, and WiFi positioning, while mmWave methods use AoA, path loss with AoA, or RSSI; RISs may improve accuracy and coverage [304].The paper describes mapping the environment and using angle and time of arrival to estimate user locations.Source: B. 6G in Simultaneous Localization and… · p. 9
  • 6G with THz is projected to enable centimeter-level accuracy in NLOS environments and network-based sensing and localization [305].Higher data rates could allow massive computations to run remotely, reducing computer-power demands in robots and autonomous driving cars.Source: B. 6G in Simultaneous Localization and… · pp. 9–10
  • THz could support gas, air-quality, health, and body sensing, while augmenting computer vision with NLOS views for rescue and sensing.The paper envisions subcentimeter positioning and 3D environmental maps without calibration or prior knowledge, as well as 6G and AI supporting future connectivity demands.Source: B. 6G in Simultaneous Localization and… · p. 10

VI. CONCLUSION

  • The survey introduces Lidar- and camera-based SLAM through sensor types and brands, then traces its development with open-source solutions.Source: §VI · p. 10
  • Its Lidar coverage includes 2D and 3D SLAM, deep learning, and challenges and future directions; its visual coverage includes camera sensors, open-source system categories, visual-inertial odometry, and deep learning.The conclusion notes visual SLAM's prominence amid deep-learning growth despite Lidar sensor stability.Source: §VI · p. 10
  • The paper identifies Lidar–vision fusion as a balanced approach to reliability and versatility.Source: §VI · p. 10
  • The paper points to challenges and opportunities across SLAM applications, future research directions for sensor-based SLAM, and a 6G vision for radio-based SLAM.Source: §VI · p. 10
Loading 1909.05214v4…