HS-SLAM: A Fast and Hybrid Strategy-Based SLAM Approach for Low-Speed Autonomous Driving

Bingxiang Kang; Guofa Li; Jie Li; Jie Zeng; Jie Zou; Kan Wang; Pengwei Zhang

REVIEW 5 major objections 5 minor 36 references

HS-SLAM: A Fast and Hybrid Strategy-Based SLAM Approach for Low-Speed Autonomous Driving

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A SLAM front end that predicts motion from the IMU, refines it with coarse-to-fine photometric alignment, and skips descriptor extraction on non-keyframes reports higher EuRoC accuracy than ORB-SLAM3 on four of five sequences with 15%…

desk verdict The speedup is plausible and consistently measured, but the accuracy-superiority claim is contradicted by the paper's own tables; with the claim narrowed, this is a solid engineering increment worth reviewing. read the letter →

arxiv 2505.20906 v1 pith:AAEIPQC6 submitted 2025-05-27 cs.RO

Bingxiang Kang , Jie Zou , Guofa Li , Pengwei Zhang , Jie Zeng , Kan Wang , Jie Li This is my paper

classification cs.RO

keywords visual-inertialSLAMhybriddirect-featuretrackingIMUposepredictionphotometricrefinementORB-SLAM3EuRoCMAVdatasetkeyframeselectionlow-speedautonomousdriving

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that a feature-based visual-inertial SLAM system can be made faster without losing accuracy by replacing the constant-velocity tracking model with a hybrid front end. In HS-SLAM, the IMU predicts the pose of each new frame, a four-layer direct photometric method refines that prediction from coarse to fine, and descriptor extraction is skipped for every frame that is not selected as a keyframe. On the EuRoC MAV benchmark, the authors report lower RMSE ATE than ORB-SLAM3 on four of five sequences and a 15.13% average reduction in per-frame tracking time. The paper's stated motivation is that such savings matter for low-speed autonomous vehicles and small embedded devices.

What carries the argument

The carrying mechanism is the two-stage fusion tracker. Stage one computes $\hat{T}_t = T_{t-1} T_{t-1\to t}^{IMU}$ from the IMU's predicted relative motion; stage two minimizes the photometric error $\varepsilon = I_{t-1}(p_i^{t-1}) - I_t(T_t p_i^t)$ on an equally scaled four-layer image pyramid, starting at the top layer and refining downward. The pyramid turns large pixel displacements at the bottom layer into small displacements at the top layer, widening the basin of convergence of the nonlinear optimization. Around this tracker, keyframe selection by pose change, average parallax, and minimum tracked points decides when descriptors are computed, so non-keyframes are handled by optical flow or by the direct refinement without descriptor extraction.

What would settle it

Run HS-SLAM on a low-speed driving sequence with ground-truth poses and disable the four-layer refinement: if trajectory accuracy and tracking time are unchanged, the refinement is not what carries the speed-accuracy claim. A sharper test is to inject increasing IMU bias into the prior $\hat{T}_t = T_{t-1} T_{t-1\to t}^{IMU}$ and record the fraction of frames where the photometric optimization fails to reduce reprojection error; the claim breaks at the bias level where that failure rate makes mean ATE exceed ORB-SLAM3's.

Watch

Extended reading notes

Core claim

The central claim is that the constant-velocity motion prior is replaceable by an IMU-predicted relative pose followed by coarse-to-fine photometric refinement, and this replacement pays for itself because only keyframes need ORB descriptors. In the reported EuRoC experiments, the hybrid system produces RMSE ATE of 0.0209 m on MH01, 0.0396 m on MH02, and 0.0367 m on MH03, compared with ORB-SLAM3's 0.0259, 0.0758, and 0.0437 m, while average tracking time per frame falls from roughly 22.5 ms to 18.9 ms. On MH04 and V102 the accuracy comparison flips, with HS-SLAM reporting 0.0694 m versus 0.0661 m and 0.0198 m versus 0.0153 m, so the paper's summary claim is that it wins in four of five sequences and loses slightly in one.

Load-bearing premise

The load-bearing premise is that for every non-keyframe the IMU-predicted relative pose lands inside the convergence basin of the four-layer photometric refinement, so the coarse-to-fine optimization finds the true pose rather than a local minimum; the paper validates this only on indoor EuRoC sequences, and its own conclusion concedes the direct method assumes grayscale invariance and can fail under abrupt illumination changes.

Editorial extensions

If this is right

Embedded visual-inertial systems can track at 50 to 65 frames per second on EuRoC-class sequences instead of 40 to 50 frames per second while keeping ORB-SLAM3's back-end mapping and loop closure.
Descriptor computation and matching become keyframe-only operations, so the per-frame cost of ordinary frames approaches the cost of optical flow tracking.
The four-layer pyramid can handle larger inter-frame motion than a single-level direct method, because coarse layers shrink apparent pixel displacements before fine layers refine them.
The size of the speedup varies by sequence, from 27.89% on MH01 and 30.08% on MH02 down to 1.59% on V102, so the practical gain depends on how many frames the keyframe criteria send to the descriptor path.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper's own evaluation, the 15% figure is an indoor-drone number; low-speed ground vehicles bring vibration, wheel slip, and outdoor lighting that change both IMU noise and the grayscale-invariance assumption.
Beyond the paper's own evaluation, the two-stage tracker is localized in the front end, so the same IMU-plus-pyramid refinement could be tried under a wheel-odometry prior or with another feature-based back end.
Beyond the paper's own evaluation, reporting how often relocalization and global photometric recovery fire would show whether skipping non-keyframe descriptors shifts failure modes elsewhere in the system.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The thing to know about HS-SLAM is that the speed result looks real and the accuracy claim does not hold up as stated. The paper takes ORB-SLAM3 and replaces the constant-velocity tracking model with an IMU-predicted relative pose, refined by a four-layer pyramid direct photometric alignment, and skips descriptor extraction for non-keyframes. That is a sensible hybrid arrangement and, as far as I can tell, not exactly present in the prior work it cites. The reported per-frame tracking time improves on every one of the five EuRoC sequences, from 1.6% to 30.1%, with an average of about 15%. The mechanism is credible and the comparison against ORB-SLAM3, the codebase it builds on, is honest.

The problem is the abstract's claim of "higher localization accuracies." The paper's own Table 2 shows HS-SLAM worse than ORB-SLAM3 on MH04 (-5.00% ATE) and V102 (-29.41% ATE), and Table 3 shows it much worse on V102 RPE (-69.62%). The text says the method is superior in four sequences, but the tables show better ATE in only three of five. That is not a minor wording issue; it is a load-bearing mismatch between the evidence and the headline. The accuracy contribution should be stated as comparable, or better on most sequences, not simply "higher." With that narrowing, the paper is still a decent engineering result: a consistent speedup at roughly equal accuracy is useful for embedded and low-speed platforms.

Other soft spots are secondary but worth noting. There is no code release, no ablation that isolates the contribution of the IMU initialization or the pyramid refinement, and the key thresholds (tau_T, tau_p, N_min, epsilon, pyramid settings) are never reported. The experiments were repeated five times, but only single RMSE values are given; there is no run-to-run variance or significance test, only the within-trajectory S.D. Those are fixable. The more substantive concern is generalization: despite the title, all evaluation is on EuRoC drone sequences. The IMU-predicted initialization could behave differently under low-speed driving with different vibration and motion characteristics, and the paper does not validate that at all.

This is not a paper that overturns anything, but it is a straightforward engineering modification with a plausible and consistently measured speed gain. If the authors release code, provide an ablation, report parameters, and correct the accuracy claim, it would be a worthwhile contribution for people working on resource-constrained visual-inertial SLAM. My vote: send it to peer review, but require the claim revision and the missing evidence before acceptance.

Referee Report

5 major / 5 minor

Summary. The paper proposes HS-SLAM, a stereo visual-inertial SLAM system built on ORB-SLAM3 that replaces the constant-velocity motion model with an IMU-predicted pose refined by a four-layer direct photometric optimization, and reduces computation by skipping descriptor extraction for non-keyframes. The authors evaluate on five EuRoC MAV sequences and report a 15.13% average tracking-time improvement while claiming higher localization accuracy than ORB-SLAM3.

Significance. The hybrid feature-direct strategy and the consistently positive speed improvement across all five sequences (Table 4) are useful engineering contributions for resource-constrained robotics. The paper provides a clear rationale for bypassing descriptor extraction on non-keyframes and demonstrates the speed gain in a standard benchmark. However, the accuracy-superiority claim is contradicted by the reported ATE/RPE results, and the stated applicability to low-speed autonomous driving is not supported by the chosen evaluation scenario. The paper does not mention code release, which limits reproducibility of the claimed improvements.

major comments (5)

[Abstract and Section 4.2] The abstract states that HS-SLAM 'achieves higher localization accuracies than ORB-SLAM3', but Table 2 shows that HS-SLAM has worse RMSE ATE on MH04 (-5.00%) and V102 (-29.41%). The statement in Section 4.2 that HS-SLAM 'demonstrates superior performance in four sequences and performs worse in one' is inconsistent with Table 2, where it is worse in two of five sequences for ATE. The accuracy-superiority claim is therefore not supported by the paper's own data and needs to be revised to a claim of comparable accuracy with sequence-dependent trade-offs.
[Table 3] The improvement percentages in the RPE comparison contain arithmetic errors. For MH01, the S.D. improvement is listed as 12.17% but the reported values 0.0083 and 0.0051 yield 38.55%. For MH02, the RMSE improvement is listed as 47.76% but 0.0537 and 0.0396 yield 26.26%, and the S.D. improvement is listed as 38.55% but 0.0533 and 0.0057 yield 89.31%. In addition, the HS-SLAM RPE S.D. for MH02 (0.0057) is inconsistent with the ATE S.D. of 0.0238 in Table 2 for the same sequence. These numeric inconsistencies undermine the quantitative accuracy summary and must be corrected.
[Title, Abstract, and Section 4.1] The paper frames HS-SLAM for low-speed autonomous driving, but all experiments use the EuRoC MAV drone dataset and are run on a laptop with an NVIDIA GTX 1660 Ti GPU. There is no evaluation on a ground vehicle, a low-speed driving scenario, or an embedded platform. The claimed applicability to low-speed autonomous driving is not demonstrated and should either be supported by additional experiments or narrowed in the title and abstract.
[Section 3.4, Eqs. (16)-(18)] The two-stage fusion method uses the IMU-predicted relative pose as an initialization for the four-layer photometric refinement, but the paper provides no ablation study separating the effect of this motion model from the effect of skipping descriptor extraction. The reader cannot determine whether the observed accuracy changes, particularly the degradation on V102, are caused by the IMU initialization or by the reduced feature set. A comparison against the original constant-velocity tracking model with descriptor extraction would clarify the source of the speed/accuracy trade-off.
[Section 4.2, Tables 2-3] The text states that experiments were repeated five times, but the tables report only a single RMSE and S.D. per sequence, where the S.D. defined in Eq. (21) is the standard deviation of per-pose errors, not the run-to-run variability. No confidence intervals or significance tests are provided, so the reported improvements cannot be assessed statistically; please clarify what the five repeats refer to and report run-to-run statistics if available.

minor comments (5)

[Table 1] The function 'EPE()' is described in Table 1 as 'eprojectionError()'; this should be 'ReprojectionError()'.
[Section 3.4] The terms 'bit-pose' and 'bit-position' are nonstandard; use 'relative pose' and 'position' instead.
[Section 3.4] The four-layer image pyramid is described as 'equally scaled' but the scale factor between layers is not specified; please state it explicitly (e.g., 0.5 per level).
[Section 3.4, Fig. 5] Figure 5 is not referenced or explained in the text; please add a sentence describing what it illustrates.
[Section 2.2] The claim that ORB is 'approximately 15 times faster than SURF' would benefit from a citation or a measurement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning: the paper's contributions are empirical engineering changes benchmarked against an external baseline, not a derivation whose outputs are presupposed by its inputs.

full rationale

HS-SLAM's central claims are (i) a speed improvement from skipping descriptor extraction on non-keyframes and (ii) comparable or better accuracy versus ORB-SLAM3 on EuRoC. Neither claim is derived by definition from a fitted input. Eq. 16 uses the IMU-predicted relative pose T_{t-1->t}^{IMU} only as an initialization, and Eqs. 17-18 refine it by minimizing photometric error; the final pose is not defined to be the IMU pose, so no self-definitional step exists. Accuracy is evaluated against the external EuRoC ground truth with EVO metrics, and speed is measured against the public ORB-SLAM3 baseline; the baseline is an external system, not a self-authored citation doing the work. References [17] and [18] are prior works by author G. Li, but they are cited only as background on learning-based visual-inertial odometry, not to justify the paper's uniqueness, correctness, or efficiency claims. The mismatch between the abstract's 'higher localization accuracies' and Tables 2-3 (HS-SLAM is worse on MH04 and V102 in ATE and on V102 in RPE) is a real evidence-claim discrepancy and a correctness risk, but it is not circularity: the data are not constructed to equal the claim. The conclusion's admitted limitation of gray-scale invariance under illumination changes also weakens generalization but does not indicate circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. Its burden comes from several unreported design parameters and strong domain assumptions about photometric consistency and IMU initialization quality. The free parameters are especially important because keyframe selection directly controls the measured speed gain, and the paper does not report or ablate them.

free parameters (5)

tau_T (pose-change threshold) = not reported
Controls keyframe selection in Eq. (13); directly trades tracking speed against map density and accuracy.
tau_p (average parallax threshold) = not reported
Controls keyframe selection in Eq. (14); affects when new keyframes are created and how much descriptor extraction is skipped.
N_min (minimum valid matches threshold) = not reported
Used in Algorithm 1 and Eq. (15) to trigger keyframe creation and tracking-failure recovery; strongly influences robustness and speed.
pyramid levels and scaling = four-layer equally scaled pyramid
The coarse-to-fine direct optimization depends on the chosen pyramid architecture, but no ablation shows sensitivity to this choice.
epsilon (descriptor matching threshold) = not reported
Defines valid descriptor matches in Eq. (11); affects match quality and the number of tracked features.

assumptions (5)

domain assumption Gray-level invariance holds between consecutive frames for the photometric error in Eq. (18).
The direct method assumes pixel intensities are stable across views; this is stated in Section 3.3 and can fail under lighting changes.
domain assumption The IMU-predicted relative pose in Eq. (16) is an accurate enough initialization for direct refinement.
The whole front-end depends on coarse matching starting from the IMU prediction; Section 3.4 does not analyze failure modes or bias drift.
domain assumption The EuRoC ground-truth trajectories are accurate and the EVO metrics are computed correctly.
Section 4.1 relies on motion-capture and laser-tracker references; the paper assumes these references are error-free enough for the claimed comparisons.
domain assumption The ORB-SLAM3 baseline was configured fairly and not disadvantaged in the comparison.
Section 4.2 compares against ORB-SLAM3 without reporting its parameter settings, so equal-tuning cannot be verified.
standard math Standard epipolar geometry, SVD pose decomposition, and triangulation are valid for initialization.
Section 3.1.1 uses the essential matrix, SVD, and triangulation without proving these standard results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HS-SLAM: A Fast and Hybrid Strategy-Based SLAM Approach for Low-Speed Autonomous Driving." pith.science (2026). https://pith.science/paper/AAEIPQC6

@misc{pith2026250520906,
  author       = {Pith},
  title        = {Pith review of: HS-SLAM: A Fast and Hybrid Strategy-Based SLAM Approach for Low-Speed Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AAEIPQC6}},
  note         = {Machine review of arXiv:2505.20906}
}

read the original abstract

Visual-inertial simultaneous localization and mapping (SLAM) is a key module of robotics and low-speed autonomous vehicles, which is usually limited by the high computation burden for practical applications. To this end, an innovative strategy-based hybrid framework HS-SLAM is proposed to integrate the advantages of direct and feature-based methods for fast computation without decreasing the performance. It first estimates the relative positions of consecutive frames using IMU pose estimation within the tracking thread. Then, it refines these estimates through a multi-layer direct method, which progressively corrects the relative pose from coarse to fine, ultimately achieving accurate corner-based feature matching. This approach serves as an alternative to the conventional constant-velocity tracking model. By selectively bypassing descriptor extraction for non-critical frames, HS-SLAM significantly improves the tracking speed. Experimental evaluations on the EuRoC MAV dataset demonstrate that HS-SLAM achieves higher localization accuracies than ORB-SLAM3 while improving the average tracking efficiency by 15%.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages

[1]

Thrun, W

S. Thrun, W. Burgard, and D. Fox, Probabilistic Robotics (Intelligent Robotics and Autonomous Agents). Cambridge, MA, USA: MIT Press, 2005

work page 2005
[2]

Simultaneous localization and mapping: Part I,

H. Durrant -Whyte and T. Bailey, “Simultaneous localization and mapping: Part I,” IEEE Robot. Autom. Mag., vol. 13, no. 2, pp. 99–110, Jun. 2006

work page 2006
[3]

A review of visual SLAM methods for autonomous driving vehicles,

J. Cheng, L. Zhang, Q. Chen, X. Hu, and J. Cai, “A review of visual SLAM methods for autonomous driving vehicles,” Eng. Appl. Artif. Intell., vol. 114, Sep. 2022, Art. no. 104992

work page 2022
[4]

A consistent parallel estimation framework for visual -inertial SLAM,

Z. Huai and G. Huang, “A consistent parallel estimation framework for visual -inertial SLAM,” IEEE Trans. Robot. , vol. 40, pp. 3734 –3755, Jul. 2024

work page 2024
[5]

Edge-Assisted Multi-Robot Visual-Inertial SLAM With Efficient Communication,

X. Liu, S. Wen, J. Zhao, T. Z. Qiu, and H. Zhang, “Edge-Assisted Multi-Robot Visual-Inertial SLAM With Efficient Communication,” IEEE Trans. Autom. Sci. Eng., vol. 22, pp. 2186–2198, 2025

work page 2025
[6]

PIPO-SLAM: Lightweight Visual -Inertial SLAM With Preintegration Merging Theory and Pose -Only Descriptions of Multiple View Geometry,

Y. Ge, L. Zhang, Y. Wu, and D. Hu, “PIPO-SLAM: Lightweight Visual -Inertial SLAM With Preintegration Merging Theory and Pose -Only Descriptions of Multiple View Geometry,” IEEE Trans. Robotics, vol. 40, pp. 2046–2059, Feb. 2024

work page 2024
[7]

PE -VINS: Accurate Monocular Visual - Inertial SLAM With Point -Edge Features,

C. Liu, H. Yu, P. Cheng, W. Sun, J. Civera, and X. Chen, “PE -VINS: Accurate Monocular Visual - Inertial SLAM With Point -Edge Features,” IEEE Trans. Intell. Vehicles, pp. 1–11, Jun. 2024

work page 2024
[8]

A survey of state-of-the-art on visual SLAM,

I. A. Kazerouni, L. Fitzgerald, G. Dooly, and D. Toal, “A survey of state-of-the-art on visual SLAM,” Expert Syst. Appl., vol. 205, p. 117734, Nov. 2022

work page 2022

Show all 36 references

[9]

LiDAR, IMU, and camera fusion for simultaneous localization and mapping: a systematic review,

Z. Fan, L. Zhang, X. Wang, Y. Shen, and F. Deng, “LiDAR, IMU, and camera fusion for simultaneous localization and mapping: a systematic review,” Artif. Intell. Rev., vol. 58, no. 6, p. 174, Mar. 2025

2025
[10]

ORB -SLAM2: An open-source SLAM system for monocular, stereo, and RGB-D cameras,

R. Mur-Artal and J. D. Tardós, “ORB -SLAM2: An open-source SLAM system for monocular, stereo, and RGB-D cameras,” IEEE Trans. Robot., vol. 33, no. 5, pp. 1255–1262, Oct. 2017

2017
[11]

Equivariant filter (EqF),

P. van Goor, T. Hamel, and R. Mahony, “Equivariant filter (EqF),” IEEE Trans. Autom. Control, vol. 68, no. 6, pp. 3501–3512, Jun. 2023

2023
[12]

EqVIO: An equivariant filter for visual -inertial odometry,

P. van Goor and R. Mahony, “EqVIO: An equivariant filter for visual -inertial odometry,” IEEE Trans. Robot., vol. 39, no. 5, pp. 3567 –3585, Oct. 2023

2023
[13]

Monocular Visual -Inertial Odometry with Planar Regularities,

C. Chen, P. Geneva, Y. Peng, W. Lee, and G. Huang, “Monocular Visual -Inertial Odometry with Planar Regularities,” in 2023 IEEE Int. Conf. Robotics Autom. (ICRA), London, United Kingdom, 2023, pp. 6224–6231

2023
[14]

VINS -Mono: A Robust and Versatile Monocular Visual -Inertial State Estimator,

T. Qin, P. Li, and S. Shen, “VINS -Mono: A Robust and Versatile Monocular Visual -Inertial State Estimator,” IEEE Trans. Robotics, vol. 34, no. 4, pp. 1004–1020, Aug. 2018

2018
[15]

ORB -SLAM3: An Accurate Open -Source Library for Visual, Visual – Inertial, and Multimap SLAM,

C. Campos, R. Elvira, J. J. G. Rodrí guez, J. M. M. Montiel, and J. D. Tardós, “ORB -SLAM3: An Accurate Open -Source Library for Visual, Visual – Inertial, and Multimap SLAM,” IEEE Trans. Robotics, vol. 37, no. 6, pp. 1874–1890, Dec. 2021

2021
[16]

DM-VIO: Delayed Marginalization Visual -Inertial Odometry,

L. v. Stumberg and D. Cremers, “DM-VIO: Delayed Marginalization Visual -Inertial Odometry,” IEEE Robotics Autom. Lett., vol. 7, no. 2, pp. 1408–1415, Apr. 2022

2022
[17]

VONAS: Network Design in Visual Odometry using Neural Architecture Search,

X. Cai, L. Zhang, C. Li, G. Li, and T. H. Li, “VONAS: Network Design in Visual Odometry using Neural Architecture Search,” in Proc. 28th ACM Int. Conf. Multimedia (MM '20), Seattle, WA, USA, 2020, pp. 727–735

2020
[18]

ATVIO: Attention Guided Visual-Inertial Odometry,

L. Liu, G. Li, and T. H. Li, “ATVIO: Attention Guided Visual-Inertial Odometry,” in 2021 IEEE Int. Conf. Acoustics, Speech and Signal Processing (ICASSP), Toronto, ON, Canada, 2021, pp. 4125 – 4129

2021
[19]

MonoSLAM: Real -Time Single Camera SLAM,

A. J. Davison, I. D. Reid, N. D. Molton, and O. Stasse, “MonoSLAM: Real -Time Single Camera SLAM,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 29, no. 6, pp. 1052–1067, Jun. 2007

2007
[20]

ORB-SLAM: A Versatile and Accurate Monocular SLAM System,

R. Mur-Artal, J. M. M. Montiel, and J. D. Tardó s, “ORB-SLAM: A Versatile and Accurate Monocular SLAM System,” IEEE Trans. Robotics, vol. 31, no. 5, pp. 1147–1163, Oct. 2015

2015
[21]

UPLP-SLAM: Unified point -line-plane feature fusion for RGB -D visual SLAM,

H. Yang, J. Yuan, Y. Gao, X. Sun, and X. Zhang, “UPLP-SLAM: Unified point -line-plane feature fusion for RGB -D visual SLAM,” Inf. Fusion, vol. 96, pp. 51–65, 2023

2023
[22]

EGLT -SLAM: Real -Time Visual-Inertial SLAM Based on Entropy -Guided Line Tracking,

X. Jia, Y. Ning, D. Chai, J. Fan, Z. Yang, X. Xi, F. Zhu, and W. Wang, “EGLT -SLAM: Real -Time Visual-Inertial SLAM Based on Entropy -Guided Line Tracking,” IEEE Sens. J., vol. 24, no. 20, pp. 32757–32771, Oct. 2024

2024
[23]

LIFT-SLAM: A deep -learning feature -based monocular visual SLAM method,

H. M. S. Bruno and E. L. Colombini, “LIFT-SLAM: A deep -learning feature -based monocular visual SLAM method,” Neurocomputing, vol. 455, pp. 97– 110, 2021

2021
[24]

Faster R -CNN Learning-Based Semantic Filter for Geometry Estimation and Its Application in vSLAM Systems,

C. Shao, L. Zhang, and W. Pan, “Faster R -CNN Learning-Based Semantic Filter for Geometry Estimation and Its Application in vSLAM Systems,” IEEE Trans. Intell. Transp. Syst., vol. 23, no. 6, pp. 5257–5266, Jun. 2022

2022
[25]

DTAM: Dense tracking and mapping in real-time,

R. A. Newcombe, S. J. Lovegrove, and A. J. Davison, “DTAM: Dense tracking and mapping in real-time,” in 2011 Int. Conf. Computer Vision (ICCV), Barcelona, Spain, 2011, pp. 2320–2327

2011
[26]

LSD-SLAM: Large-Scale Direct Monocular SLAM,

J. Engel, T. Schöps, and D. Cremers, “LSD-SLAM: Large-Scale Direct Monocular SLAM,” in Computer Vision – ECCV 2014, D. Fleet, T. Pajdla, B. Schiele, and T. Tuytelaars, Eds., Cham, Switzerland: Springer, 2014, pp. 834–849

2014
[27]

Direct Sparse Odometry,

J. Engel, V. Koltun, and D. Cremers, “Direct Sparse Odometry,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 40, no. 3, pp. 611–625, Mar. 2018

2018
[28]

SVO: Fast semi -direct monocular visual odometry,

C. Forster, M. Pizzoli, and D. Scaramuzza, “SVO: Fast semi -direct monocular visual odometry,” in 2014 IEEE Int. Conf. Robotics Autom. (ICRA), Hong Kong, China, 2014, pp. 15–22

2014
[29]

Keyframe -based visual –inertial odometry using nonlinear optimization,

S. Leutenegger, S. Lynen, M. Bosse, R. Siegwart, and P. Furgale, “Keyframe -based visual –inertial odometry using nonlinear optimization,” Int. J. Robotics Res., vol. 34, no. 3, pp. 314–334, 2014

2014
[30]

Learned Uncertainty Calibration for Visual Inertial Localization,

S. Tsuei, S. Soatto, P. Tabuada, and M. B. Milam, “Learned Uncertainty Calibration for Visual Inertial Localization,” in 2021 IEEE Int. Conf. Robotics Autom. (ICRA), Xi'an, China, 2021, pp. 5311–5317

2021
[31]

Dynam-SLAM: An Accurate, Robust Stereo Visual-Inertial SLAM Method in Dynamic Environments,

H. Yin, S. Li, Y. Tao, J. Guo, and B. Huang, “Dynam-SLAM: An Accurate, Robust Stereo Visual-Inertial SLAM Method in Dynamic Environments,” IEEE Trans. Robotics, vol. 39, no. 1, pp. 289–308, Feb. 2023

2023
[32]

Ground - Fusion: A Low-cost Ground SLAM System Robust to Corner Cases,

J. Yin, A. Li, W. Xi, W. Yu, and D. Zou, “Ground - Fusion: A Low-cost Ground SLAM System Robust to Corner Cases,” in 2024 IEEE Int. Conf. Robotics Autom. (ICRA), Yokohama, Japan, 2024, pp. 8603– 8609

2024
[33]

Distinctive Image Features from Scale-Invariant Keypoints,

D. G. Lowe, “Distinctive Image Features from Scale-Invariant Keypoints,” Int. J. Computer Vision, vol. 60, no. 2, pp. 91–110, Nov. 2004

2004
[34]

Speeded-Up Robust Features (SURF),

H. Bay, A. Ess, T. Tuytelaars, and L. Van Gool, “Speeded-Up Robust Features (SURF),” Computer Vision and Image Understanding, vol. 110, no. 3, pp. 346–359, 2008

2008
[35]

ORB: An efficient alternative to SIFT or SURF,

E. Rublee, V. Rabaud, K. Konolige, and G. Bradski, "ORB: An efficient alternative to SIFT or SURF," in 2011 Int. Conf. Comput. Vis. (ICCV), Barcelona, Spain, 2011, pp. 2564–2571

2011
[36]

The EuRoC micro aerial vehicle datasets

Burri M, Nikolic J, Gohl P, et al. The EuRoC micro aerial vehicle datasets. Int J Robot Res. vol. 35, no. 10, pp. 1157–1163, 2016

2016

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Thrun, W

S. Thrun, W. Burgard, and D. Fox, Probabilistic Robotics (Intelligent Robotics and Autonomous Agents). Cambridge, MA, USA: MIT Press, 2005

work page 2005

[2] [2]

Simultaneous localization and mapping: Part I,

H. Durrant -Whyte and T. Bailey, “Simultaneous localization and mapping: Part I,” IEEE Robot. Autom. Mag., vol. 13, no. 2, pp. 99–110, Jun. 2006

work page 2006

[3] [3]

A review of visual SLAM methods for autonomous driving vehicles,

J. Cheng, L. Zhang, Q. Chen, X. Hu, and J. Cai, “A review of visual SLAM methods for autonomous driving vehicles,” Eng. Appl. Artif. Intell., vol. 114, Sep. 2022, Art. no. 104992

work page 2022

[4] [4]

A consistent parallel estimation framework for visual -inertial SLAM,

Z. Huai and G. Huang, “A consistent parallel estimation framework for visual -inertial SLAM,” IEEE Trans. Robot. , vol. 40, pp. 3734 –3755, Jul. 2024

work page 2024

[5] [5]

Edge-Assisted Multi-Robot Visual-Inertial SLAM With Efficient Communication,

X. Liu, S. Wen, J. Zhao, T. Z. Qiu, and H. Zhang, “Edge-Assisted Multi-Robot Visual-Inertial SLAM With Efficient Communication,” IEEE Trans. Autom. Sci. Eng., vol. 22, pp. 2186–2198, 2025

work page 2025

[6] [6]

PIPO-SLAM: Lightweight Visual -Inertial SLAM With Preintegration Merging Theory and Pose -Only Descriptions of Multiple View Geometry,

Y. Ge, L. Zhang, Y. Wu, and D. Hu, “PIPO-SLAM: Lightweight Visual -Inertial SLAM With Preintegration Merging Theory and Pose -Only Descriptions of Multiple View Geometry,” IEEE Trans. Robotics, vol. 40, pp. 2046–2059, Feb. 2024

work page 2024

[7] [7]

PE -VINS: Accurate Monocular Visual - Inertial SLAM With Point -Edge Features,

C. Liu, H. Yu, P. Cheng, W. Sun, J. Civera, and X. Chen, “PE -VINS: Accurate Monocular Visual - Inertial SLAM With Point -Edge Features,” IEEE Trans. Intell. Vehicles, pp. 1–11, Jun. 2024

work page 2024

[8] [8]

A survey of state-of-the-art on visual SLAM,

I. A. Kazerouni, L. Fitzgerald, G. Dooly, and D. Toal, “A survey of state-of-the-art on visual SLAM,” Expert Syst. Appl., vol. 205, p. 117734, Nov. 2022

work page 2022