Visual SLAM: What are the Current Trends and What to Expect?

I. INTRODUCTION

• Direct methods estimate motion from pixel-level data, while indirect methods extract and match feature points; each has different strengths and trade-offs.Indirect methods are described as precise and robust to photometric changes, while direct methods use all camera pixels but can pose large-scale optimization challenges.Source: §I · pp. 1–2
• The survey reviews forty-five VSLAM papers and classifies them by diverse characteristics, including their objectives and approaches.It aims to provide a reference for researchers working to improve VSLAM techniques.Source: §I · p. 2

II. EVOLUTION OF VISUAL SLAM ALGORITHMS

• PTAM divided tracking and mapping into parallel threads to reduce computational cost and support real-time performance.It also introduced bundle adjustment to jointly optimize camera poses and the 3D map.Source: §II · p. 2
• Early systems illustrate trade-offs: Mono-SLAM lacked global optimization and loop closure, while DTAM produced detailed maps at high computational cost,.Mono-SLAM’s maps contained landmarks without further area detail.Source: §II · pp. 2–3
• Later approaches included hybrid SVO, which combines feature-based and direct methods, and LSD-SLAM, which supports large-scale maps, global optimization, and loop closure,.SVO lacks loop closure and global optimization; LSD-SLAM has a challenging, computationally intensive initialization,.Source: §II · p. 3
• ORB-SLAM 2.0 added parallel tracking, local mapping, and loop-closing threads, but its maps have unknown scale and it struggles in texture-less or repetitive environments.ORB-SLAM 3.0 supports monocular, RGB-D, and stereo cameras and provides improved pose estimates.Source: §II · p. 3

III. RELATED SURVEYS

• Prior surveys variously covered sensor classes, traditional and semantic methods, graph optimization and deep learning, or feature-based approaches, with stated limits in scope or generalizability –.Other work examined situational awareness in SLAM and VSLAM applications [35].Source: §III · p. 4
• Topic-focused surveys addressed deep learning, loop closure, driverless vehicles, autonomous driving, and operation in dynamic or rough environments, –.Their scopes included deep learning in VO and loop closure detection.Source: §III · p. 4
• This survey categorizes recent VSLAM publications by contributions, criteria, and research objectives, analyzes current trends, and introduces potential contributions for researchers.Source: §III · p. 4

A. Sensors and Data Acquisition

• Monocular, stereo, and RGB-D cameras provide different depth and color information, but can struggle under low illumination or high dynamic range.Stereo cameras use multiple image sensors to perceive depth; RGB-D cameras provide depth and color.Source: A. Sensors and Data Acquisition · pp. 4–5
• Event cameras report pixel-level brightness changes, supporting high-speed motion and wide dynamic range, but provide insufficient information at low motion rates.Their unsynchronized output also challenges traditional vision algorithms.Source: A. Sensors and Data Acquisition · p. 5
• Multi-camera overlap can help address occlusion, camouflage, sensor failure, and sparse trackable texture, while camera-only systems may still face blur, mismatches, and dynamic-object challenges.The survey also identifies event/frame fusion and lidar or IMU integration as existing solutions.Source: A. Sensors and Data Acquisition · p. 5

B. Target Environments

• Traditional VSLAM often presumes a static world, while unexpected changes such as moving objects can substantially degrade state-estimation quality.Dynamic-scene systems may detect motion with optical flow or RANSAC and exclude moving objects as outliers.Source: B. Target Environments · p. 5
• Off-road environments have fewer trackable points than urban areas, increasing risks of localization and loop-closure failure; performance may differ across indoor and outdoor settings.Off-road scenes may include moving vegetation, clouds, and weak texture, unlike the structural landmarks common in urban areas.Source: B. Target Environments · p. 5

C. Visual Features Processing

• Indirect VSLAM requires feature extraction for pose estimation, and CNNs are also used to extract image features for visual odometry, pose estimation, and loop closure.ORB offers fast extraction and matching without a large accuracy loss compared with SIFT and SURF.Source: C. Visual Features Processing · p. 5

D. System Evaluation

• Many VSLAM studies use public datasets, including RAWSEEDS, SceneNet RGB-D, TUM RGB-D, and NTU VIRAL, with differing sensor and ground-truth contents.NTU VIRAL includes data from a UAV equipped with lidars, cameras, IMUs, and multiple UWBs.Source: D. System Evaluation · p. 5
• EuRoC MAV, KITTI, TartanAir, and ICL-NUIM provide additional evaluation settings, while the Event Camera Dataset supports high-speed event-camera evaluation.The Event Camera Dataset includes inertial measurements and intensity images captured by a motion-capture system.Source: D. System Evaluation · p. 6

E. Semantic Level

• Semantic information supplements geometry-based data with conceptual scene knowledge, including through pre-trained object-recognition modules and CNNs.The survey presents this as a direction used in recent VSLAM works.Source: E. Semantic Level · p. 6
• Semantic VSLAM combines tracking, local mapping, loop closing, and Non-Rigid Context Culling to handle pose, map, and temporal-object processing.NRCC masks or segments unstable instances such as people, reducing feature points to process; using semantics without greatly increasing computational cost remains a challenge.Source: E. Semantic Level · p. 6

V. VSLAM APPROACHES BASED ON THE MAIN OBJECTIVES

• The survey selected highly cited recent publications from Google Scholar, Scopus, and DBLP, then categorized relevant papers by their main objectives.It also examined papers cited by the selected publications.Source: §V · p. 6

A. Objective I: Multi-sensor Processing

• CoSLAM combines independently moving cameras across platforms to reconstruct robust maps from overlapping views in dynamic settings, with increased hardware and computational costs.It combines intra- and inter-camera pose estimation and mapping, and uses KLT feature tracking.Source: 1) Employing Multiple Cameras · pp. 6–7
• Yang et al. use independent cameras, ORB features, and CNN-based loop-closure matching for off-road conditions including occlusion and texture sparsity.Their experiment used a dataset produced by a panoramic camera and integrated navigation system.Source: 1) Employing Multiple Cameras · p. 7
• MultiCol-SLAM supports multiple fisheye cameras through keyframe-based ORB-SLAM processing and multi-camera loop closure, but requires substantial computing power.Its simultaneous threads support real-time operation.Source: 1) Employing Multiple Cameras · p. 7
• CamVox combines RGB-D camera outputs with Livox lidar and an IMU, and reports real-time operation in real-world robot tests.Its contribution includes autonomous lidar-camera calibration in uncontrolled environments.Source: 2) Employing Multiple Sensors · p. 7
• VIRAL SLAM couples camera, lidar, IMU, and UWB with lidar-map-based visual-feature marginalization, but is computationally intensive.It was tested in simulation and on the generated NTU VIRAL dataset.Source: 2) Employing Multiple Sensors · p. 7
• Ultimate SLAM combines event-camera data, standard frames, and IMU for high-speed positioning, avoiding motion-blur problems but facing event/frame synchronization challenges.The authors report performance on the Event Camera Dataset and tests on an autonomous quadrotor.Source: 2) Employing Multiple Sensors · p. 7
• Nguyen et al. combine monocular-camera features and UWB landmarks in an ORB-SLAM-based indirect method that operates under UWB multipath effects.The system was tested on a generated dataset with hand-carried movements simulating an aerial robot; camera-UWB synchronization was addressed using a timestamped camera pose for each image.Source: 2) Employing Multiple Sensors · pp. 7–8

Employing Lines/Points Data

• Structural lines encode dominant directions and global orientation for camera-pose estimation; StructSLAM uses EKF in low-feature and featureless conditions.It was evaluated on RAWSEEDS 2009 indoor scenes and generated sequential images.Source: 1) Employing Lines/Points Data · p. 9
• PL-SLAM fuses points and lines for pose estimation in non-dynamic, low-texture settings, but incurs computational cost and may benefit from additional geometric primitives.The authors tested it on a generated dataset and TUM RGB-D.Source: 1) Employing Lines/Points Data · p. 9
• PL-SLAM tracks points and line segments across stereo frames and outperformed stereo ORB-SLAM 2.0, with feature-tracking time as a drawback.It was evaluated on EuRoC and KITTI.Source: 1) Employing Lines/Points Data · p. 9
• Lim et al. combine optical-flow line tracking with degeneracy avoidance for monocular point-and-line VSLAM, but lack adaptive optimization-parameter selection.They tested the method on EuRoC MAV.Source: 1) Employing Lines/Points Data · p. 9

Using Extra Features

• DQV-SLAM uses progressive Bayes updates to avoid linearizing nonlinear spatial transformations and uses ORB features for data association in dynamic scenes.The method produced reliable experimental results on KITTI and EuRoC, but is computationally demanding and lacks a probabilistic interpretation for stochastic pose modeling.Source: 2) Using Extra Features · p. 9
• SPM-SLAM resolves pose ambiguity when at least two artificial planar markers are visible, but requires numerous markers and its dynamic-scene handling was not measured.The authors examined it using marker sequences from two rooms joined by a door.Source: 2) Using Extra Features · p. 9

Deep Learning

• LIFT-SLAM combines learned feature descriptors with geometry and outperformed conventional feature-based and deep-learning VSLAM systems in accuracy on KITTI and EuRoC instances.Its computationally intensive pipeline and unoptimized CNN design result in near-real-time performance.Source: 3) Deep Learning · p. 9
• Naveed et al. used deep reinforcement learning trained on realistic simulators for route planning, reporting generalization across actual indoor and outdoor environments.The approach used ORB-SLAM as its base system and was evaluated with navigation episodes in challenging, texture-less environments.Source: 3) Deep Learning · p. 9

Dynamic Environments

• DS-SLAM uses optical flow and SegNet to exclude dynamic objects before pose estimation, while constructing a dense semantic map through five system threads.It was tested in real-world settings and on TUM RGB-D; the survey notes segmentation limitations and computationally intensive features.Source: 1) Dynamic Environments · p. 10
• SOF-SLAM combines semantic segmentation with semantic optical flow to skip changing features and outperformed ORB-SLAM 2.0 in extremely dynamic TUM RGB-D and real-world tests.Its non-static feature recognition is ineffective and relies on only two consecutive frames.Source: 1) Dynamic Environments · p. 10
• Cheng et al. use optical flow to remove dynamic feature points and report good performance in dynamic indoor TUM RGB-D tests; their offline motion-analysis threshold complicates use across varied dynamic settings.The system is based on the ORB-SLAM pipeline.Source: 1) Dynamic Environments · p. 10
• Yang et al. combine semantic segmentation, motion consistency, and geometric restrictions to create static semantic maps, with evaluations on Oxford and TUM RGB-D.The method improved location accuracy, but can encounter problems in corridors or information-poor places.Source: 1) Dynamic Environments · p. 10

Deep Learning-based Solutions

• DXSLAM uses HF-NET-derived local and global descriptions and offline Bag of Words for loop closure, and outperformed ORB-SLAM 2.0 and DS-SLAM in reported accuracy.It runs in real time without a GPU on contemporary CPUs, but has a complex feature-extraction architecture.Source: 2) Deep Learning-based Solutions · p. 10
• A self-supervised multitask CNN outputs fixed-length 256-bit feature codes and supports dense-map creation with monocular and RGB-D cameras.The system uses tracking, local-mapping, and loop-closing threads and was tested on TUM and two Kinect corridor and office datasets.Source: 2) Deep Learning-based Solutions · p. 10
• A UAV-oriented system uses indoor-trained depth estimation with ORB-SLAM 2.0 and demonstrated enhanced pose accuracy, but struggles without texture and needs CPU and GPU resources for real-time performance.Its CNN was trained on more than 48,000 indoor examples and was evaluated on TUM RGB-D and with a real-world drone.Source: 2) Deep Learning-based Solutions · p. 11

Using Artificial Landmarks

• UcoSLAM combines natural landmarks and fiducial markers, supports keypoint-only, marker-only, and mixed modes, and can operate without tags or features.Its multiple-thread execution makes it time-consuming.Source: 3) Using Artificial Landmarks · p. 11

Devices with Limited Processing Capabilities

• EdgeSLAM distributes fine-grained semantic VSLAM modules between mobile devices and an edge server, adapting parameters to bandwidth and latency to avoid duplicate segmentation processing.It was evaluated on monocular TUM RGB-D, KITTI, and a created dataset.Source: 1) Devices with Limited Processing… · p. 11
• ProSLAM is a lightweight, single-threaded stereo framework with triangulation, motion estimation, map management, and re-localization modules, and yielded robust KITTI and EuRoC results.The survey notes weaknesses in rotation estimation and the absence of bundle adjustment.Source: 1) Devices with Limited Processing… · p. 11
• VPS-SLAM combines geometry, object detection, and visual or visual-inertial odometry to build sparse semantic maps and predict robot states, with results on par with well-known VSLAM methods.Its semantic mapping is limited to a small number of object types, and tests included TUM RGB-D indoor instances and an aerial robot.Source: 1) Devices with Limited Processing… · p. 11
• A low-cost indoor system based on OpenVSLAM uses EPnP and RANSAC for pose estimation and delivered accurate results under difficult lighting in building tests.The authors also proposed estimating the number of frames and visual elements needed for reasonable localization accuracy.Source: 1) Devices with Limited Processing… · pp. 11–12

Computation Offloading

• Edge-SLAM keeps tracking on the robot and offloads remaining ORB-SLAM 2.0 modules to the edge, maintaining local and global maps without sacrificing accuracy under fewer resources.Its decoupled architecture is complex, and the system is limited to short-term settings; performance may degrade over multiple days.Source: 2) Computation Offloading · p. 12

Versatility

• OpenVSLAM is an adaptable open-source framework compatible with monocular, stereo, and RGB-D cameras, and can store or reuse reconstructed maps.The survey reports better tracking accuracy and efficiency than ORB-SLAM and ORB-SLAM 2.0, but says its code was discontinued over concerns about similarities infringing ORB-SLAM 2.0 rights.Source: E. Objective V: Versatility · p. 12
• OV2SLAM combines direct and indirect methods and limits feature extraction to keyframes to reduce computational load; it surpassed several popular techniques in performance and accuracy.It was evaluated on EuRoC, KITTI, and TartanAir across indoor and outdoor experiments.Source: E. Objective V: Versatility · p. 12
• DROID-SLAM supports monocular, stereo, and RGB-D cameras, operates in real time, and reports greater accuracy and robustness than well-known monocular and stereo tracking methods.It was tested on TartanAir, EuRoC, and TUM RGB-D, and its monocular training allows stereo and RGB-D use without retraining.Source: E. Objective V: Versatility · p. 12
• iRotate uses active observation coverage to reduce the robot's travel distance while mapping, with results on par with cutting-edge VSLAM methods in simulation and on a real robot.The robot may face start-stop cases when local paths are replanned.Source: E. Objective V: Versatility · p. 12

• VSLAM approaches in this category aim to estimate the robot’s position and orientation as accurately as possible.Source: F. Objective VI: Visual Odometry · p. 12
• Dynamic-SLAM uses CNN-based moving-object detection, selective tracking, and missed-feature correction to improve pose estimation, but has high computational costs and risks dynamic/static misclassification.Source: 1) Deep Neural Networks · p. 12
• Code-SLAM19 condenses dense scene geometry using monocular intensity images and a depth auto-encoder, with promising accuracy and performance in indoor EuRoC examples.Source: 1) Deep Neural Networks · pp. 12–13
• DeepVO estimates monocular poses from color-frame sequences, while a neural-graph alternative adds loop-closure correction and outperformed DeepVO on 2D and 3D mazes.DeepVO combines FlowNet optical flow with LSTM layers; the neural-graph method requires connection to a SLAM framework for re-localization signals.Source: 1) Deep Neural Networks · p. 13
• DeepFactors jointly optimizes pose and depth for dense monocular mapping, combining learned and model-based components but requiring GPUs for real-time performance.It was evaluated on ICL-NUIM and TUM RGB-D after training on roughly 1.4 million ScanNet images.Source: 1) Deep Neural Networks · p. 13
• A dense RGB-D SLAM approach uses non-redundant keyframes, entropy-based loop-closure detection, and drift avoidance, but its loop closure and keyframe selection still need improvement.Source: 2) In-depth Adjacent Frame Processing · p. 13
• Other approaches remove moving objects with Bayesian keypoint probabilities, reuse leader traces for indoor navigation, or build 3D text maps; their scope includes sparse-SLAM, generated-data, and text-availability constraints.DP-SLAM has high computational costs; TextSLAM also faces short-letter interpretation and large text-dictionary challenges.Source: 2) In-depth Adjacent Frame Processing · p. 13, 3) Various Feature Processing · pp. 13–14
• An OGM-based modified ORB-SLAM supports real-time route planning in GPS-denied conditions, while plane-based CPA-SLAM uses RGB-D tracking and graph optimization but supports only planes.Source: 3) Various Feature Processing · p. 14

• Improving visual odometry is the most common objective among the surveyed VSLAM applications, with pose estimation and real-world viability also prominent.Source: A. Statistics · p. 14
• TUM RGB-D is the most-used evaluation dataset, followed by EuRoC MAV and KITTI, while many studies also test on author-generated datasets.The survey suggests generated datasets are used to demonstrate real-world operation and end-to-end application potential.Source: A. Statistics · p. 14
• Most reviewed papers do not use semantic data, which the survey associates with semantic-processing costs and the possibility that incorrect scene information adds noise.Geometry-based VSLAM is often designed as plug-and-play, using camera data for localization and mapping with little effort.Source: A. Statistics · p. 14
• More than half of the approaches can work in dynamic environments, while most are tested indoors or in both indoor and outdoor settings.The survey cautions that methods developed under restricted assumptions may not retain the same accuracy elsewhere.Source: A. Statistics · p. 14

B. Analyzing the Current Trends

• Deep learning is a prominent direction for feature extraction, visual odometry, loop closure, object detection, and semantic segmentation, though limited object classes can risk misclassification and pose error.Source: B. Analyzing the Current Trends · pp. 14–15
• Dense maps capture high-dimensional scene information but are computationally demanding in real time, whereas sparse maps cost less but can omit needed information.Frame losses during peak processing can also harm performance, and tightly coupled modules make changes to one component consequential for others.Source: B. Analyzing the Current Trends · p. 15
• Semantic scene information is a potential input to pose estimation, trajectory planning, and loop-closure detection.Source: B. Analyzing the Current Trends · p. 15
• Loop-closure detection can add latency and cost as reconstructed maps grow, while texture-poor scenes can cause pose drift; the survey identifies optimization and complementary scene cues as research opportunities.Source: B. Analyzing the Current Trends · pp. 15–16

VII. CONCLUSIONS

• The survey categorizes recent VSLAM work by experimental environment, novelty, detection and tracking algorithms, semantic viability, and performance, and discusses contributions, drawbacks, trends, and open issues.Source: §VII · p. 16