Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Short egocentric video clips can forecast whole-environment motion maps up to 20 seconds ahead.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

EgoMoD predicts full-environment Maps of Dynamics (flow, direction, entropy) from an 8-frame egocentric video plus robot pose, trained with privileged global supervision.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A new task formulation with a plausible method, but the missing pose-only baseline means the paper hasn't yet shown the video is doing the work. the 4 major comments →

arxiv 2603.00167 v2 pith:7IZ3YFE5 submitted 2026-02-26 cs.RO

EgoMoD: Predicting Global Maps of Dynamics from Local Egocentric Observations

classification cs.RO
keywords maps of dynamicsegocentric visionglobal motion predictionrobot navigationbird's-eye view predictionprivileged learningvideo foundation modelscrowd dynamics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 introduces EgoMoD, a method that learns to predict future global Maps of Dynamics—flow, dominant direction, and directional entropy over the entire environment—from just a short clip of egocentric video and the robot's pose. It argues that local visual cues, combined with position, contain enough information to infer aggregate motion patterns in areas the robot has never seen, replacing the external global sensing infrastructure that prior map-of-dynamics methods require. If correct, a robot navigating a hospital corridor could anticipate congestion and plan preemptively using only its onboard camera and localization. Experiments in a simulated hospital and a photorealistic replica of an office show accurate forecasts and zero-shot transfer to real images.

Core claim

The central claim is that global, future motion tendencies can be predicted directly from local egocentric observations. EgoMoD processes an 8-frame video clip with a frozen video foundation model to extract spatio-temporal features, fuses these with a learned pose embedding through a transformer attention module, and decodes the result into bird's-eye-view maps of flow, dominant direction, and directional entropy for the full environment at a future horizon T (10 or 20 s). Training uses privileged supervision: ground-truth maps built from external cameras observing the whole scene. At inference the external sensing is gone; only the onboard camera and pose are used, and the model must forec

What carries the argument

The key mechanism is a pose-conditioned video-to-map architecture. A frozen video foundation model encodes short clips into patch-level spatio-temporal features; a pose token is prepended and processed through a transformer self-attention block, letting every visual patch attend to the robot's global position; a convolutional decoder then upsamples the fused representation into three allocentric map heads (flow magnitude, dominant direction, directional entropy). This design converts local visual dynamics into allocentric predictions, and the pose token prevents the network from memorizing pose-to-map correspondences while enforcing that visual cues are grounded in the world frame.

Load-bearing premise

The training assumes that a short local video clip at time t contains enough information to predict aggregate motion over the whole environment for the next 10-20 seconds, i.e., that future global dynamics are sufficiently determined by the current local visual context plus the environment's stationary motion patterns.

What would settle it

A controlled experiment where a robot observes a corridor with perfectly uniform flow while another area changes its motion pattern (e.g., a door opens and redirects flow) within the 10-20 s horizon; if the model cannot anticipate the change, it would show that local cues alone are insufficient. Alternatively, a quantitative evaluation on a new environment layout without retraining would reveal whether the learned mapping transfers beyond the training site.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Robots can anticipate congestion and plan proactive routes using standard onboard cameras and localization, without CCTV or multi-robot sensing.
  • Map of dynamics approaches no longer require long observation periods at deployment; a short clip at the current time replaces historical accumulation.
  • The approach can be zero-shot transferred from simulation to real sensor images when motion patterns are matched.
  • The learned maps provide flow, direction, and entropy that could be used as priors in trajectory forecasting and social navigation stacks.
  • The architecture allows forecasting over the whole environment, including areas completely outside the field of view.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The dependency on a privileged global supervisor during training could be replaced by self-supervised consistency between local predictions and slow-moving global statistics, potentially removing the need for external cameras.
  • Because the model is environment-specific, a natural extension is cross-environment transfer by conditioning on static floor plan embeddings or object layout, which the paper does not explore.
  • The 8-frame clip and 10-20 s horizon imply a kind of short-term stationarity; if crowd flows shift on sub-minute timescales, the upper bound on usable horizon is set by how long the local visual context remains predictive.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes EgoMoD, a supervised architecture that takes a short egocentric RGB video clip (8 frames) plus robot pose and predicts a global, allocentric Map of Dynamics (MoD) — flow magnitude, dominant direction, and directional entropy — over a future horizon T (10 or 20 s), with n << T. Training uses privileged MoDs computed from external global observations; at inference only onboard camera and localization are used. The model uses a frozen V-JEPA video encoder, a pose token fused through a Transformer, and three BEV decoder heads with class-imbalance weighting and gradient-structure losses. Experiments are conducted in a simulated hospital, comparing against a local observation-only baseline (Local-StefMap), plus a qualitative deployment in a photorealistic replica of the authors' office and a physical office with matched motion patterns. The paper claims to be the first approach to learn future MoD prediction directly from short egocentric video.

Significance. If the central claim is established, EgoMoD would be a practically relevant step: it would replace external global sensing at inference with standard onboard sensing for MoD-based navigation. The architecture is reasonable and the training-supervision scheme is clearly described. The paper is also commendably transparent about its limitations (environment specificity, absence of suitable public datasets, fixed horizon). However, the current experiments do not isolate the mechanism that the claim rests on — namely, that the egocentric video adds predictive signal beyond what pose and the environment layout already determine. The empirical validation is therefore not yet sufficient to support the headline contribution, though the deficiencies are addressable with controlled ablations and additional evaluation.

major comments (4)
  1. [§V-E, Table III; §IV-A]
  2. [§V-C, §V-E, Table III]
  3. [§V-E]
  4. [§V-F, §VI, Abstract]
minor comments (5)
  1. [§III-A and §IV]
  2. [§III-B, Eq. (2)]
  3. [§V-B, Table II]
  4. [§V-F, Fig. 7]
  5. [General]

Circularity Check

0 steps flagged

No circularity found: EgoMoD's forecast targets are constructed from a disjoint future interval, not from its local inputs or fitted constants.

full rationale

The paper's derivation chain is not circular. Target MoDs are constructed by accumulating detections over a T-second interval (Sec. III-C.2, Eqs. 1-3), while the model inputs are an 8-frame egocentric clip and pose from an earlier short interval n with n << T; the paper explicitly states that 'the supervision-MoDs are aggregated over T seconds of future observations, our model only uses egocentric frames for an initial short interval n as input'. No parameter is fitted to the test MoDs, and the test scene is held out (Sec. V-C), so the reported predictions are not statistical fits of the evaluation targets. The baseline Local-StefMap is weaker, but that is an evaluation-strength concern, not a definitional reduction. The Discussion candidly acknowledges that EgoMoD is environment-specific and that the real-world validation uses a photorealistic replica with matched motion patterns; this limits generalization claims but does not make the predicted maps equivalent to the inputs by construction. No load-bearing self-citations, no imported uniqueness theorems, and no fitted quantity renamed as a prediction were found. The absence of a pose-only or static-prior ablation is a legitimate experimental gap, but a missing control is not a demonstrated circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The ledger is dominated by domain assumptions about the environment, the privileged supervision, and the sufficiency of local visual cues. The learned network weights are the main fitted quantity, and several hand-chosen hyperparameters (map resolution, direction bins, horizon, loss weights) shape the supervision and the objective. No new physical or invented entities are postulated.

free parameters (7)
  • Learned EgoMoD network weights (attention module + decoder heads) = Trained on 2,000 hospital sequences (80/20 split)
    The central mapping from video+pose to MoDs is a fitted function, not a derivation.
  • Direction bins B = 8
    Discretizes [0,2π) for orientation histograms; affects direction and entropy labels (Table I).
  • Grid cell size = 0.30 m
    Sets the spatial resolution of the MoD maps and the density of supervision labels (Table I).
  • Prediction horizon T = 10 / 20 s
    Defines the future aggregation window used to build supervision and the evaluation horizon (Table I).
  • Input clip length n = 8 frames
    The claim of predicting from 'short egocentric video' depends on this specific clip length (§IV).
  • Valid/background loss weights (w_valid, w_bg) = 5.0 / 0.95
    Hand-chosen weighting to counter class imbalance between motion and static cells (Table I).
  • Gradient loss weight λ_grad and Huber threshold β = 1.0 / 0.1
    Hand-chosen loss hyperparameters that affect the training objective (Table I).
axioms (6)
  • domain assumption MoDs built by accumulating orientation histograms over T seconds faithfully represent future motion tendencies.
    Section III-B defines flow, direction, and entropy from aggregated detections; this aggregation is the supervised target and the evaluation ground truth.
  • domain assumption Future aggregate motion over T can be inferred from an n << T egocentric clip because motion patterns are recurrent and structured.
    Section III-C.2 states that the setup 'ensures that the network learns to infer future dynamics from a short initial clip'.
  • domain assumption Privileged global observations are available during training for each deployment environment.
    Section III-C.2 assumes CCTV or other robots provide global detections; the limitation section admits the model is environment-specific and requires such data.
  • domain assumption PedSim-based simulated agents with added queuing behaviors are representative of real human dynamics.
    Section V-A describes the hospital simulation, and Section V-F relies on matching real-office motion patterns to the simulation.
  • domain assumption Frozen V-JEPA2 features contain sufficient spatio-temporal cues for motion direction and flow.
    Section IV uses the frozen V-JEPA2 backbone as the sole visual encoder; no fine-tuning is performed.
  • domain assumption Nearest-neighbor centroid association between consecutive frames is sufficient to estimate motion directions for MoD construction.
    Section III-B.2 explicitly states that sophisticated data association or tracking is not required.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of EgoMoD: Predicting Global Maps of Dynamics from Local Egocentric Observations." pith.science (2026). https://pith.science/paper/7IZ3YFE5

@misc{pith2026260300167,
  author       = {Pith},
  title        = {Pith review of: EgoMoD: Predicting Global Maps of Dynamics from Local Egocentric Observations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7IZ3YFE5}},
  note         = {Machine review of arXiv:2603.00167}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Efficient navigation in dynamic environments requires anticipating how motion patterns evolve beyond the robot's immediate perceptual range, enabling preemptive rather than purely reactive planning in crowded scenes. Maps of Dynamics (MoDs) offer a structured representation of motion tendencies in space useful for long-term global planning, but constructing them traditionally requires global environment observations over extended periods of time. We introduce EgoMoD, the first approach that learns to predict future MoDs directly from short egocentric video clips collected during robot operation. Our method learns to infer environment-wide motion tendencies from local dynamic cues using a video- and pose-conditioned architecture trained with MoDs computed from external observations as privileged supervision, allowing local observations to serve as predictive signals of global motion structure. Thanks to this, we offer the capacity to forecast future motion dynamics over the whole environment rather than merely extend past patterns in the robot's field of view. As a site-specific dynamic prior, EgoMoD replaces the external global sensing infrastructure required by prior MoD methods at inference time with standard onboard sensors. Experiments in large simulated environments show that EgoMoD predicts future MoDs under limited observability, while evaluation with real images showcases its zero-shot transferability to real systems.

Figures

Figures reproduced from arXiv: 2603.00167 by David Morilla-Cabello, Eduardo Montijano, Iacopo Catalano, Jorge Pena-Queralta.

Figure 1
Figure 1. Figure 1: EgoMoD learns to predict global maps of dynamics (bottom) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: EgoMoD architecture overview. A short egocentric video [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Example of construction of our Maps of Dynamics [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the hospital simulation environment shown from [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Photorealistic office simulated environment [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results showing predicted flow, entropy, and direction maps from our model on test scenes, compared to ground truth. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative results for the MoD predicted by our network on [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Rheos: Modelling Continuous Motion Dynamics in Hierarchical 3D Scene Graphs

    cs.RO 2026-03 conditional novelty 6.0

    Rheos embeds online semi-wrapped Gaussian mixture models of directional motion into 3D scene graph navigational nodes and outperforms discrete histogram baselines on continuous and discrete metrics.

Reference graph

Works this paper leans on

35 extracted references · 1 linked inside Pith · cited by 1 Pith paper

  1. [1]

    A survey on socially aware robot navigation: Taxonomy and future challenges,

    P. T. Singamaneni, P. Bachiller-Burgos, L. J. Manso, A. Garrell, A. San- feliu, A. Spalanzani, and R. Alami, “A survey on socially aware robot navigation: Taxonomy and future challenges,”The Int. J. of Robotics Research, vol. 43, no. 10, pp. 1533–1572, 2024

  2. [2]

    Survey of maps of dynamics for mobile robots,

    T. P. Kucner, M. Magnusson, S. Mghames, L. Palmieri, F. Verdoja, C. S. Swaminathan, T. Krajn ´ık, E. Schaffernicht, N. Bellotto, M. Hanheide, et al., “Survey of maps of dynamics for mobile robots,”The Int. J. of Robotics Research, vol. 42, no. 11, pp. 977–1006, 2023

  3. [3]

    Enabling flow awareness for mobile robots in partially observable environments,

    T. P. Kucner, M. Magnusson, E. Schaffernicht, V . H. Bennetts, and A. J. Lilienthal, “Enabling flow awareness for mobile robots in partially observable environments,”IEEE Robotics & Automation L., vol. 2, no. 2, pp. 1093–1100, 2017

  4. [4]

    Spectral analysis for long-term robotic mapping,

    T. Krajnik, J. P. Fentanes, G. Cielniak, C. Dondrup, and T. Duckett, “Spectral analysis for long-term robotic mapping,” inIEEE Int. Conf. on Robot. Autom., pp. 3706–3711, 2014

  5. [5]

    Robotic exploration for learning human motion patterns,

    S. Molina, G. Cielniak, and T. Duckett, “Robotic exploration for learning human motion patterns,”IEEE Trans. on Robotics, vol. 38, no. 2, pp. 1304–1318, 2021

  6. [6]

    V-jepa 2: Self- supervised video models enable understanding, prediction and planning,

    M. Assran, A. Bardes, D. Fan, Q. Garrido, R. Howes, M. Muckley, A. Rizvi, C. Roberts, K. Sinha, A. Zholus,et al., “V-jepa 2: Self- supervised video models enable understanding, prediction and planning,” arXiv preprint arXiv:2506.09985, 2025

  7. [7]

    Fremen: Frequency map enhancement for long-term mobile robot autonomy in changing environments,

    T. Krajn ´ık, J. P. Fentanes, J. M. Santos, and T. Duckett, “Fremen: Frequency map enhancement for long-term mobile robot autonomy in changing environments,”IEEE Trans. on Robotics, vol. 33, no. 4, pp. 964–977, 2017

  8. [8]

    Time- varying pedestrian flow models for service robots,

    T. Vintr, S. Molina, R. Senanayake, G. Broughton, Z. Yan, J. Ulrich, T. P. Kucner, C. S. Swaminathan, F. Majer, M. Stachov ´a,et al., “Time- varying pedestrian flow models for service robots,” inEuropean Conf. on Mobile Robots, pp. 1–7, 2019

  9. [9]

    Learning state-space models for mapping spatial motion patterns,

    J. Shi and T. P. Kucner, “Learning state-space models for mapping spatial motion patterns,” inEuropean Conf. on Mobile Robots, pp. 1–6, 2023

  10. [10]

    Learning temporal maps of dynamics for mobile robots,

    J. Shi and T. P. Kucner, “Learning temporal maps of dynamics for mobile robots,”Robotics & Autonomous Syst., vol. 184, p. 104853, 2025

  11. [11]

    Neural implicit flow fields for spatio-temporal motion mapping,

    Y . Zhu, S.-M. Yang, A. Rudenko, T. P. Kucner, A. J. Lilienthal, and M. Magnusson, “Neural implicit flow fields for spatio-temporal motion mapping,” inInt. Conf. on Learning Representations, 2026

  12. [12]

    Non-parametric modeling of spatio- temporal human activity based on mobile robot observations,

    M. Stuede and M. Schappler, “Non-parametric modeling of spatio- temporal human activity based on mobile robot observations,” in IEEE/RSJ Int. Conf. on Intell. Robots and Syst., pp. 126–133, 2022

  13. [13]

    Bayesian floor field: Transferring people flow predictions across environments,

    F. Verdoja, T. P. Kucner, and V . Kyrki, “Bayesian floor field: Transferring people flow predictions across environments,” inIEEE/RSJ Int. Conf. on Intell. Robots and Syst., pp. 12801–12807, 2024

  14. [14]

    Stef-lstm: A hybrid framework inte- grating periodic and sequential modeling for mapping motion patterns,

    Z. Yan, J. Shi, and T. P. Kucner, “Stef-lstm: A hybrid framework inte- grating periodic and sequential modeling for mapping motion patterns,” inEuropean Conf. on Mobile Robots, pp. 1–8, 2025

  15. [15]

    Event-triggered maps of dynamics: A framework for modeling spatial motion patterns in non-stationary environments,

    J. Shi, Q. Guo, and T. P. Kucner, “Event-triggered maps of dynamics: A framework for modeling spatial motion patterns in non-stationary environments,” inIEEE/RSJ Int. Conf. on Intell. Robots and Syst., pp. 12924–12931, IEEE, 2025

  16. [16]

    Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data,

    T. Salzmann, B. Ivanovic, P. Chakravarty, and M. Pavone, “Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data,” in Eur . Conf. on Comput. Vision, pp. 683–700, Springer, 2020

  17. [17]

    Social lstm: Human trajectory prediction in crowded spaces,

    A. Alahi, K. Goel, V . Ramanathan, A. Robicquet, L. Fei-Fei, and S. Savarese, “Social lstm: Human trajectory prediction in crowded spaces,” inIEEE Conf. Comput. Vision & Pattern Recognition, pp. 961– 971, 2016

  18. [18]

    Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting,

    Y . Yuan, X. Weng, Y . Ou, and K. M. Kitani, “Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting,” inIEEE Int. Conf. on Comput. Vision, pp. 9813–9823, 2021

  19. [19]

    Stochastic trajectory prediction via motion indeterminacy diffusion,

    T. Gu, G. Chen, J. Li, C. Lin, Y . Rao, J. Zhou, and J. Lu, “Stochastic trajectory prediction via motion indeterminacy diffusion,” inIEEE Conf. Comput. Vision & Pattern Recognition, pp. 17113–17122, 2022

  20. [20]

    Sicnav-diffusion: Safe and interactive crowd navigation with diffusion trajectory predictions,

    S. Samavi, A. Lem, F. Sato, S. Chen, Q. Gu, K. Yano, A. P. Schoellig, and F. Shkurti, “Sicnav-diffusion: Safe and interactive crowd navigation with diffusion trajectory predictions,”IEEE Robotics & Automation L., 2025

  21. [21]

    Flowbot: Flow-based modeling for robot navigation,

    D. Dugas, K. Cai, O. Andersson, N. Lawrance, R. Siegwart, and J. J. Chung, “Flowbot: Flow-based modeling for robot navigation,” in IEEE/RSJ Int. Conf. on Intell. Robots and Syst., pp. 8799–8805, 2022

  22. [22]

    Trajflow: Learning distributions over trajectories for human behavior prediction,

    A. M ´esz´aros, J. F. Schumann, J. Alonso-Mora, A. Zgonnikov, and J. Kober, “Trajflow: Learning distributions over trajectories for human behavior prediction,” inIEEE Intelligent V ehicles Symposium, pp. 184– 191, IEEE, 2024

  23. [23]

    Probabilistic dynamic crowd prediction for social navigation,

    S. H. Kiss, K. Katuwandeniya, A. Alempijevic, and T. Vidal-Calleja, “Probabilistic dynamic crowd prediction for social navigation,” inIEEE Int. Conf. on Robot. Autom., pp. 9269–9275, 2021

  24. [24]

    Long-term human motion prediction using spatio-temporal maps of dynamics,

    Y . Zhu, A. Rudenko, T. P. Kucner, A. J. Lilienthal, and M. Magnusson, “Long-term human motion prediction using spatio-temporal maps of dynamics,”IEEE Robotics & Automation L., 2025

  25. [25]

    Leveraging maps of spatial motion patterns to enhance long-term adaptive trajectory prediction with diffusion models,

    J. Shi and T. P. Kucner, “Leveraging maps of spatial motion patterns to enhance long-term adaptive trajectory prediction with diffusion models,” inEuropean Conf. on Mobile Robots, pp. 1–8, 2025

  26. [26]

    Allo-centric occupancy grid prediction for urban traffic scene using video prediction networks,

    R. Asghar, L. Rummelhard, A. Spalanzani, and C. Laugier, “Allo-centric occupancy grid prediction for urban traffic scene using video prediction networks,” inInt. Conf. on Control, Automation, Robotics and Vision, pp. 255–260, 2022

  27. [27]

    Mapping high-level semantic regions in indoor environments without object recognition,

    R. Bigazzi, L. Baraldi, S. Kousik, R. Cucchiara, and M. Pavone, “Mapping high-level semantic regions in indoor environments without object recognition,” inIEEE Int. Conf. on Robot. Autom., pp. 7686–7693, 2024

  28. [28]

    Monoc- ular semantic occupancy grid mapping with convolutional variational encoder–decoder networks,

    C. Lu, M. J. G. Van De Molengraft, and G. Dubbelman, “Monoc- ular semantic occupancy grid mapping with convolutional variational encoder–decoder networks,”IEEE Robotics & Automation L., vol. 4, no. 2, pp. 445–452, 2019

  29. [29]

    Predicting semantic map representations from images using pyramid occupancy networks,

    T. Roddick and R. Cipolla, “Predicting semantic map representations from images using pyramid occupancy networks,” inIEEE Conf. Com- put. Vision & Pattern Recognition, pp. 11138–11147, 2020

  30. [30]

    Occupancy antic- ipation for efficient exploration and navigation,

    S. K. Ramakrishnan, Z. Al-Halah, and K. Grauman, “Occupancy antic- ipation for efficient exploration and navigation,” inEuropean Conf. on Mobile Robots, pp. 400–418, Springer, 2020

  31. [31]

    Fiery: Future instance prediction in bird’s- eye view from surround monocular cameras,

    A. Hu, Z. Murez, N. Mohan, S. Dudas, J. Hawke, V . Badrinarayanan, R. Cipolla, and A. Kendall, “Fiery: Future instance prediction in bird’s- eye view from surround monocular cameras,” inIEEE Int. Conf. on Comput. Vision, pp. 15273–15282, 2021

  32. [32]

    Predicting occupancy distri- butions of walking humans with convolutional neural networks,

    J. Doellinger, M. Spies, and W. Burgard, “Predicting occupancy distri- butions of walking humans with convolutional neural networks,”IEEE Robotics & Automation L., vol. 3, no. 3, pp. 1522–1528, 2018

  33. [33]

    Person tracking in large public spaces using 3-d range sensors,

    D. Br ˇsˇci´c, T. Kanda, T. Ikeda, and T. Miyashita, “Person tracking in large public spaces using 3-d range sensors,”IEEE Transactions on Human- Machine Systems, vol. 43, no. 6, pp. 522–534, 2013

  34. [34]

    Towards safe navigation through crowded dynamic environments,

    Z. Xie, P. Xin, and P. Dames, “Towards safe navigation through crowded dynamic environments,” inIEEE/RSJ Int. Conf. on Intell. Robots and Syst., pp. 4934–4940, IEEE, 2021

  35. [35]

    Ultralytics yolo11,

    G. Jocher and J. Qiu, “Ultralytics yolo11,” 2024

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.