Pith. sign in

REVIEW 4 major objections 7 minor 22 references

A comparison of extended object tracking with multi-modal sensors in indoor environment

T0 review · 4 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A stereo camera costing under 400 euros can track a mobile robot indoors about as well as a LiDAR costing more than ten times as much, according to this preliminary study.

desk verdict An honest preliminary study whose headline 'similar performance' claim rests on visual inspection, not ground truth. read the letter →

arxiv 2411.18476 v1 pith:34T2MBBU submitted 2024-11-27 cs.RO cs.CV

Jiangtao Shuai , Martin Baerveldt , Manh Nguyen-Duc , Anh Le-Tuan , Manfred Hauswirth , Danh Le-Phuoc This is my paper · ORCID
classification cs.ROcs.CV
keywords extendedobjecttrackingstereocameraLiDARpointcloudindoormobilerobotGaussianprocesstargetmodelDBSCANsmartfactory
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 reports a preliminary comparison of extended object tracking on 3D point clouds from a LiDAR sensor and a much cheaper stereo camera, both tracking a single mobile robot indoors. It claims that, using the same heuristic detector and Gaussian-process extended object tracker, the stereo camera achieves tracking performance similar to the LiDAR's, despite costing less than one-tenth as much. The tracking results are judged visually from estimated trajectories and extents in two motion scenarios, with detection rates of 98% for LiDAR and 91% for the camera. If the claim holds, surveillance and situational awareness in smart factories could be built around sub-400-euro stereo cameras instead of multi-thousand-euro LiDAR units.

What carries the argument

The method combines a heuristic detector with an extended object tracker. The detector removes ground points using RANSAC-fitted planes, restricts points to a predefined operation area, clusters remaining points with DBSCAN, computes PCA bounding boxes, and selects the target cluster by comparing geometric features against the robot's known dimensions. The selected points feed an extended object tracker that models the target's extent as a star-convex hypersurface (a shape whose boundary each radial line from the center crosses once) whose radial function is a Gaussian process, updated by an iterated extended Kalman filter with a constant-velocity motion model. This pipeline is identical for both sensors, so the comparison isolates the sensor's point-cloud quality and the fixed hyperparameters.

What would settle it

Track the robot along a known path measured by its onboard odometry or an external motion-capture system, compute per-timestep trajectory error for both sensors, and check whether the stereo camera's error is genuinely close to the LiDAR's; if it is not, the claim of similar performance fails.

Watch

Extended reading notes

Core claim

The central claim is that a stereo-camera point cloud can drive an extended object tracking pipeline as effectively as a LiDAR point cloud for single-robot tracking in a room-scale indoor setting. The paper asserts that the estimated trajectories and spatial extents produced from the Intel RealSense D435i are similar to those from the Blickfeld Cube 1 in both a turning trajectory and a straight trajectory. This is presented as evidence that the price difference between the sensors, more than tenfold, does not translate into a comparable gap in tracking behavior for this application, and that stereo-based extended object tracking is feasible.

Load-bearing premise

The paper assumes that visually similar estimated trajectories and extents are enough to prove similar tracking accuracy, because no ground-truth measurements are used and the paper explicitly defers that evaluation to future work.

Editorial extensions

If this is right

  • Room-scale single-robot tracking for smart-factory surveillance could be built on a sub-400-euro stereo camera instead of a multi-thousand-euro LiDAR, cutting sensor cost by more than an order of magnitude.
  • Because the detector is a fast heuristic rather than a deep network, the same tracking pipeline can run on edge devices.
  • The reported detection rate is 91% with the camera and 98% with the LiDAR, so the stereo pipeline occasionally loses the target for a frame even when the trajectories look similar.
  • The camera's short valid depth range (about 3 meters) confines the approach to room-scale indoor tracking rather than the long-range maritime or automotive settings where LiDAR is normally used.

Reading between the lines

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

  • Since the paper observes that stereo-camera noise grows with depth, replacing the fixed detector hyperparameters with depth-adaptive thresholds could close the 7-point detection gap and is a natural next step.
  • The cost-saving conclusion would be stronger with a quantitative benchmark against ground truth, because the current visual comparison alone may not capture small tracking errors that matter for collision avoidance.
  • The same detector-and-tracker combination could in principle transfer to other low-cost depth sensors with different noise profiles, but the paper does not test that transfer.
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.

Referee Report

4 major / 7 minor

Summary. The paper presents a preliminary comparison of extended object tracking (EOT) on 3D point clouds from a Blickfeld Cube 1 LiDAR and an Intel RealSense D435i stereo camera in an indoor environment. The authors develop a heuristic object detector that removes ground points, clusters candidates with DBSCAN, and selects the target by a feature-cost criterion; the detected points feed a Gaussian-process star-convex EOT framework with an iterated EKF. The paper reports detection rates of 98% for LiDAR and 91% for the stereo camera, shows qualitative trajectory plots in both sensors' local frames, and concludes that the stereo camera achieves tracking performance similar to that of the LiDAR at a fraction of the cost.

Significance. If the central comparison were quantitatively established, this study would be a useful data point for cost-effective indoor EOT, potentially enabling sub-400 euro stereo cameras as substitutes for multi-thousand-euro LiDAR in factory-floor tracking. The paper is transparent about its preliminary nature and explicitly lists its limitations, and it builds on well-established EOT machinery from the cited literature. However, the headline claim of 'similar performance' currently rests on qualitative visual inspection and detection-rate statistics rather than tracking-error metrics, so the significance is conditional on a quantitative evaluation being added.

major comments (4)
  1. [Section 3.3, Figure 5, and Section 4] The central claim that the stereo camera achieves similar performance to LiDAR is supported only by qualitative visual inspection of the estimated centroid trajectories and extents plotted in each sensor's own coordinate frame. The paper itself states in Section 4, 'In the future, we plan to evaluate the tracking results using ground truth data from the robot's onboard sensors,' confirming that no ground-truth accuracy assessment is currently made. Without a quantitative metric such as position RMSE, extent error, or a statistical comparison across trials, the visual overlap cannot distinguish 'similar accuracy' from 'similarly wrong estimates.' This is load-bearing because the entire contribution is the cross-sensor comparison.
  2. [Section 3.3] The two sensors have different sampling rates (4.4 Hz for LiDAR vs 30 Hz for the camera), different local coordinate frames, and different noise characteristics. Figure 5 plots trajectories in each sensor's Cartesian plane, so the apparent similarity is not evaluated on a common reference. The authors should resample and align the estimated trajectories to a common frame and report per-timestep errors with respect to a reference trajectory, or at minimum provide a quantitatively validated sensor alignment that justifies the visual comparison.
  3. [Section 3.3] The reported detection rates (98% for LiDAR and 91% for the camera) are not a substitute for tracking accuracy. A higher detection rate does not imply lower position or extent error, particularly because the detector's thresholds and feature weights (Algorithm 1, parameters 𝛾3, 𝑤𝑓, 𝑓𝑝) are hand-tuned per sensor. The paper should report tracking metrics (e.g., position RMSE, OSPA, or shape-error measures) at matched timestamps and discuss how missed detections and false positives propagate through the EKF, otherwise the cost-saving conclusion is unsupported.
  4. [Sections 3.1 and 3.2] The comparison relies on many manually selected per-sensor parameters: ground-plane thresholds 𝛾0 and 𝛾2, RANSAC parameters (𝛼, MAX_ITER, n), clustering parameters 𝑚𝑐, feature prior 𝑓𝑝, feature weights 𝑤𝑓, detection threshold 𝛾3, and the GP/EKF noise parameters. While this is not a logical circularity, it is a robustness gap: the claimed performance similarity may depend critically on these hand-tuned values. A brief sensitivity analysis over the most influential parameters (or at least over the stereo-camera noise-related parameters) would strengthen the conclusion that the comparison is not an artifact of the selected settings.
minor comments (7)
  1. [Section 2.2] The mathematical notation has inconsistent spacing and formatting, e.g., '𝑓 (𝑥𝑘+1| 𝑥𝑘)' and the state-vector definition; please proofread the math typography throughout the paper.
  2. [Section 3.2] The units for the process noise parameter are written as '𝜎𝑐 = 0.05√𝑚2𝑠−3', which does not appear to be a standard unit expression; please correct or clarify the intended units.
  3. [Section 3.1] The ground-plane coefficient vectors such as [0, 0, 1, 1]𝑇 for LiDAR and [0, 1, 0, −0.5]𝑇 for the camera imply different coordinate conventions (z-up vs y-up), but the sensor coordinate frames are not defined; please state the conventions explicitly.
  4. [Section 3.3] The text says the robot was tracked 'within a range of 3 meters' and cites the D435i's operational range as 0.3 to 3 meters; please clarify whether the entire trajectory lies inside the validated stereo depth range and how the LiDAR's larger range affects the comparison.
  5. [Figure 5] The caption refers to 'the light blue curve' but the text mentions 'light blue curve' without consistent color naming; please ensure the color references exactly match the rendered figure.
  6. [Figure 5] The captions (a) 'Camera one', (b) 'Lidar one', (c) 'Camera two', (d) 'Lidar two' are inconsistently styled; please unify the naming (e.g., 'Camera, turning motion' and 'LiDAR, turning motion').
  7. [References] Several bibliographic entries are incomplete, such as [15] and [19] missing access dates or full URLs; please complete the reference list.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper is an empirical sensor comparison, and its main weakness is missing ground-truth validation, which is an evidence gap rather than a circular derivation.

full rationale

The paper's derivation chain is an experiment: sensor point clouds are filtered by a heuristic detector, then fed into an extended object tracking filter based on prior work [4]/[11], and the resulting trajectories are visually compared. No quantity is fitted to one sensor's output and then used to predict the other sensor's behavior; the tracking hyperparameters are stated fixed values (e.g., sigma_f = 0.01 m, sigma_r = 0.005 m, l = pi/6 rad) taken from prior work, and they are not tuned to the comparison outcome. The 'similar performance' conclusion is based on Figure 5, and Section 4 explicitly says ground-truth evaluation is future work; this means the claim is under-supported empirically, but it is not circular because the conclusion is not encoded in the detector cost function, the tracking equations, or the hyperparameter settings. The citation to [4], a co-authored EOT implementation, is a method reference rather than a self-citation that forecloses alternatives or supplies an unverified premise. Therefore no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim relies on the detector's hand-tuned parameters (ground plane thresholds, feature prior, weights, clustering parameters), on the star-convex GP target model, on the assumption that stereo depth error is small enough at 3 m, and on the visual comparison as a substitute for ground truth.

free parameters (6)
  • Ground plane thresholds gamma_0 and gamma_2 = 1 m and 0.02 m
    Hand-set thresholds for initial plane inlier selection and subsequent ground point removal; performance of the detector depends on them.
  • RANSAC parameters (alpha, MAX_ITER, n) = 0.1, 100, 10
    Voxel downsample rate and RANSAC iteration/inlier counts for plane estimation, chosen without systematic tuning analysis.
  • Detection feature prior f_p and weights w_f = [0.39,0.33,0.21,0.005,0.13,0.08,0.07]^T and [0.5,0.5,0.5,100,2,1,1]^T
    Hand-specified expected geometric features of the TurtleBot from URDF and weighting of feature differences; the detection cost J = w_f^T |f_box - f_p| directly selects the target.
  • Detection threshold gamma_3 and clustering parameters m_c = gamma_3=1, m_c=[0.03,30]
    Threshold on detection cost and DBSCAN eps/min_points; critical for separating robot from noise in camera point clouds.
  • Operation area bounds m_area = Camera: x in [-4,4], y in [-2,1], z in [0,2.5]; LiDAR: x in [-4,4], y in [0,2.7], z in [-4,2]
    Manual crop of the workspace; assumes target stays inside these bounds.
  • GP and EKF noise/hyperparameters = sigma_c=0.05, sigma_r=0.1, sigma_f=0.01, sigma_r=0.005, sigma_n=0.001, l=pi/6, eta_f=0.001
    Taken from prior work [4] and applied without re-tuning per sensor; the tracking result depends on these values.
assumptions (5)
  • domain assumption Ground plane is a single planar surface estimable by RANSAC and constant over the experiment.
    Used in Algorithm 1 to remove floor points; fails in non-planar or changing environments.
  • domain assumption The target's shape is adequately represented by a star-convex hypersurface parameterized by a Gaussian process.
    Adopted from [11]; all contour measurements are assumed generated by this model.
  • domain assumption Non-contour point removal retains only measurements from the target contour before EOT update.
    The EOT model assumes contour-generated measurements; the paper states 'we first remove all measurements that do not originate from the contour' (Section 3.2).
  • domain assumption Stereo camera depth error within 3 m (less than 2%) is small enough for EOT to match LiDAR.
    Underpins the claim that the D435i can substitute the Cube 1 at this range; not empirically validated with error statistics in tracking.
  • domain assumption DBSCAN clustering with fixed parameters produces one cluster corresponding to the target robot.
    The detector relies on DBSCAN with m_c; the authors note that fixed hyperparameters make it hard to filter noise when depth varies (Section 3.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of A comparison of extended object tracking with multi-modal sensors in indoor environment." pith.science (2026). https://pith.science/paper/34T2MBBU

@misc{pith2026241118476,
  author       = {Pith},
  title        = {Pith review of: A comparison of extended object tracking with multi-modal sensors in indoor environment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/34T2MBBU}},
  note         = {Machine review of arXiv:2411.18476}
}
read the original abstract

This paper presents a preliminary study of an efficient object tracking approach, comparing the performance of two different 3D point cloud sensory sources: LiDAR and stereo cameras, which have significant price differences. In this preliminary work, we focus on single object tracking. We first developed a fast heuristic object detector that utilizes prior information about the environment and target. The resulting target points are subsequently fed into an extended object tracking framework, where the target shape is parameterized using a star-convex hypersurface model. Experimental results show that our object tracking method using a stereo camera achieves performance similar to that of a LiDAR sensor, with a cost difference of more than tenfold.

Figures

Figures reproduced from arXiv: 2411.18476 by the authors.

Figure 3
Figure 3. Point cloud from D435i (a) Cube LiDAR (b) D435i camera [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Detected robot points (blue) in point clouds of two sensors [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Tracking results of two motion trajectories with Cube LiDAR and D435i camera. The coordinate in each image is on a Cartesian plane of the corresponding sensor’s local coordinates. The light blue curve in each image represents the motion trajectory of the estimated centroid. Four representative timesteps’ results are visualized in each image: the blue dashed line represents the estimated shape, the blue dot represent… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 20 canonical work pages

  1. [4]

    Baerveldt, M

    M. Baerveldt, M. E. López, E. F. Brekke, Extended target pmbm tracker with a gaussian process target model on lidar data, in: 2023 26th International Conference on Information Fusion (FUSION), IEEE, 2023, pp. 1–8

    work page 2023

  2. [11]

    Wahlström, E

    N. Wahlström, E. Özkan, Extended target tracking using gaussian processes, IEEE Transactions on Signal Processing 63 (2015) 4165–4178

    work page 2015

  3. [1]

    Granström, M

    K. Granström, M. Baum, A tutorial on multiple extended object tracking, Authorea Preprints (2023)

    work page 2023

  4. [2]

    Bewley, Z

    A. Bewley, Z. Ge, L. Ott, F. Ramos, B. Upcroft, Simple online and realtime tracking, in: 2016 IEEE international conference on image processing (ICIP), IEEE, 2016, pp. 3464–3468

    2016

  5. [3]

    Pandharipande, C.-H

    A. Pandharipande, C.-H. Cheng, J. Dauwels, S. Z. Gurbuz, J. Ibanez-Guzman, G. Li, A. Pi- azzoni, P. Wang, A. Santra, Sensing and machine learning for automotive perception: A review, IEEE Sensors Journal 23 (2023) 11097–11115

    work page 2023

  6. [5]

    X. Cao, J. Lan, X. R. Li, Y. Liu, Extended object tracking using automotive radar, in: 2018 21st International Conference on Information Fusion (FUSION), IEEE, 2018, pp. 1–5

    work page 2018

  7. [6]

    Y. Xia, P. Wang, K. Berntorp, L. Svensson, K. Granström, H. Mansour, P. Boufounos, P. V. Orlik, Learning-based extended object tracking using hierarchical truncation measurement model with automotive radar, IEEE Journal of Selected Topics in Signal Processing 15 (2021) 1013–1029

    work page 2021

  8. [7]

    K. A. Ruud, E. F. Brekke, J. Eidsvik, Lidar extended object tracking of a maritime vessel using an ellipsoidal contour model, in: 2018 Sensor Data Fusion: Trends, Solutions, Applications (SDF), IEEE, 2018, pp. 1–6

    work page 2018

Show all 22 references
  1. [8]

    Ester, H.-P

    M. Ester, H.-P. Kriegel, J. Sander, X. Xu, et al., A density-based algorithm for discovering clusters in large spatial databases with noise, in: kdd, volume 96, 1996, pp. 226–231

    1996

  2. [9]

    GmbH, Blickfeld cube 1, 2024

    B. GmbH, Blickfeld cube 1, 2024. URL: https://www.blickfeld.com/lidar-sensor-products/ cube-1/, accessed: 2024-07-05

    2024

  3. [10]

    Corporation, Intel realsense d435i, 2024

    I. Corporation, Intel realsense d435i, 2024. URL: https://www.intelrealsense.com/ depth-camera-d435i/, accessed: 2024-07-05

    2024

  4. [12]

    Y. Guo, H. Wang, Q. Hu, H. Liu, L. Liu, M. Bennamoun, Deep learning for 3d point clouds: A survey, IEEE transactions on pattern analysis and machine intelligence 43 (2020) 4338–4364

    2020

  5. [13]

    W. Liu, J. Sun, W. Li, T. Hu, P. Wang, Deep learning on point clouds and its application: A survey, Sensors 19 (2019) 4188

    2019

  6. [14]

    Kelly, B

    C. Kelly, B. Wilkinson, A. Abd-Elrahman, O. Cordero, H. A. Lassiter, Accuracy assessment of low-cost lidar scanners: An analysis of the velodyne hdl–32e and livox mid–40’s temporal stability, Remote Sensing 14 (2022) 4220

    2022

  7. [15]

    URL: https://www.stereolabs

    StereoLabs, Stereolabs docs: Depth sensing overview, 2024. URL: https://www.stereolabs. com/docs/depth-sensing

    2024

  8. [16]

    Granstrom, M

    K. Granstrom, M. Baum, S. Reuter, Extended object tracking: Introduction, overview and applications, arXiv preprint arXiv:1604.00970 (2016)

    2016 arXiv

  9. [17]

    K. G. Derpanis, Overview of the ransac algorithm, Image Rochester NY 4 (2010) 2–3

    2010

  10. [18]

    Dimitrov, C

    D. Dimitrov, C. Knauer, K. Kriegel, G. Rote, Bounds on the quality of the pca bounding boxes, Computational Geometry 42 (2009) 772–789

    2009

  11. [19]

    Sucan, J

    I. Sucan, J. Kay, urdf-ros wiki, URL http://wiki. ros. org/urdf/. Data de consulta 14 (2019)

    2019

  12. [20]

    URL: http://wiki.ros.org/Robots/TurtleBot4, accessed: 2024-07-05

    Turtlebot4, 2024. URL: http://wiki.ros.org/Robots/TurtleBot4, accessed: 2024-07-05

    2024

  13. [21]

    El Yabroudi, K

    M. El Yabroudi, K. Awedat, R. C. Chabaan, O. Abudayyeh, I. Abdel-Qader, Adaptive dbscan lidar point cloud clustering for autonomous driving applications, in: 2022 IEEE International Conference on Electro Information Technology (eIT), IEEE, 2022, pp. 221–224

    2022

  14. [22]

    J. Lin, A. Puthiyavinayagam, S. Liu, M. Kurowski, J.-J. Gehrt, R. Zweigel, D. Abel, Real-time multi-object tracking using adaptive filtering and filter banks for maritime applications, in: 2021 European Control Conference (ECC), IEEE, 2021, pp. 2239–2244

    2021

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.