Pith. sign in

REVIEW 2 major objections 5 minor 148 references

Camera-based AI keeps pedestrians and cyclists safe only when four vision tasks—detection, tracking, trajectory prediction, and intent recognition—operate as one pipeline.

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 →

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

T0 review · deepseek-v4-flash

2026-08-04 13:24 UTC pith:B4RKQX7W

load-bearing objection A solid, clearly organized survey of four VRU vision tasks, but the advertised LLM coverage is absent from the body—a real overclaim, not a fatal flaw. the 2 major comments →

arxiv 2510.03314 v2 pith:B4RKQX7W submitted 2025-09-30 cs.CV cs.AI

From Camera-Based Sensing to Reasoning: A Comprehensive Review Toward Proactive Vulnerable Road User Safety

Shucheng Zhang , Yan Shi , Bingzhang Wang , Yuang Zhang , Muhammad Monjurul Karim , Kehua Chen , Chenxi Liu , Mehrdad Nasri
show 1 more author
Yinhai Wang
This is my paper
classification cs.CV cs.AI
keywords vulnerable road userscamera-based sensingpedestrian safetyobject detectionmulti-object trackingtrajectory predictionintent recognitiondeep learning
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.

This review argues that camera-based artificial intelligence for protecting pedestrians, cyclists, and other vulnerable road users has outgrown the detection-only view that dominates earlier surveys. It organizes the last five years of research into four connected tasks: detect and classify, track and re-identify, predict trajectories, and recognize intent, claiming that together these form the backbone of proactive, collision-avoiding safety systems. It also names four obstacles that block real-world deployment: scarce and biased data, poor generalization to new environments, edge-computing limits, and hardware and weather degradation. If the framing holds, future research and system design should be structured around this integrated pipeline rather than isolated tasks, and emerging AI paradigms such as vision transformers, large language models, and diffusion models should be adopted for uncertainty modeling and semantic reasoning.

Core claim

The paper's central claim is that proactive VRU safety through cameras is best understood as a hierarchical pipeline: visual perception (detection and classification), motion modeling (tracking and re-identification, then trajectory prediction), and behavior understanding (intent recognition and reasoning). The authors review recent methods in each stage and assert that these stages, when integrated, enable early risk anticipation and timely intervention. They further claim that four major challenges—data scarcity and bias, limited generalization, edge-aware model optimization, and hardware and environmental constraints—currently prevent reliable, equitable, and scalable deployment. The surv

What carries the argument

The organizing device is the four-task pipeline: detection and classification, tracking and re-identification, trajectory prediction, and intent recognition and prediction. Each task feeds the next, from localizing a VRU to inferring its intention within the traffic context. The paper also relies on a VRU taxonomy—pedestrians, cyclists, non-motorized device users, wheelchair users, and scooter or skateboard users—to ground the sensing challenges that each task must address.

Load-bearing premise

The load-bearing premise is that the four-task decomposition—detection, tracking, trajectory prediction, and intent recognition—is the correct and sufficient organizing structure for proactive VRU safety, but the survey offers no empirical comparison showing this pipeline outperforms end-to-end or task-agnostic alternatives.

What would settle it

A controlled study comparing a four-task pipeline against a single end-to-end vision model on the same camera streams, measuring near-miss prediction accuracy and false-alarm rates, would settle the matter: if the end-to-end model matches or beats the explicit pipeline without the task hierarchy, the paper's central organizing claim is undercut.

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

If this is right

  • Future VRU safety systems will be expected to integrate all four tasks rather than optimize detection alone.
  • Benchmarks should measure multi-task performance—tracking consistency, forecast accuracy, and intent correctness—not just detection metrics.
  • Vision transformers, large language models, and diffusion models will be adopted for representation learning, uncertainty modeling, and semantic reasoning in safety pipelines.
  • Data collection and augmentation efforts should target underrepresented VRU classes and mitigate demographic and geographic bias.
  • Deployment-ready systems will require model compression and edge optimization alongside robustness to adverse weather and sensor degradation.

Where Pith is reading between the lines

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

  • If the four-task pipeline is the right decomposition, then joint multi-task training across the stages could yield efficiency and accuracy gains over independently trained modules; a testable extension is whether joint training outperforms staged pipelines on real-world near-miss prediction.
  • The review suggests intent recognition is the least mature stage; an editor's inference is that closing the intent gap will matter more than further detection improvements for proactive safety.
  • The four challenges are not independent: synthetic data generation and simulation could simultaneously alleviate data scarcity, class imbalance, and generalization limitations, so research should treat them as one coupled problem.
  • The framing invites a system-level benchmark that evaluates end-to-end safety outcomes—such as reductions in predicted collisions or false alarms—rather than per-task accuracy, which would provide a stronger test of the pipeline's value.

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

2 major / 5 minor

Summary. The paper is a survey of camera-based AI methods for vulnerable road user (VRU) safety. It proposes a four-task pipeline — detection and classification, tracking and re-identification, trajectory prediction, and intent recognition and prediction — as the backbone of proactive VRU protection, and reviews representative deep-learning methods for each task across Sections III-A through III-D. It then identifies four practical challenges (data scarcity/bias, generalization, edge deployment, hardware/environmental limits) and outlines future directions in Section IV. The paper contains no experimental validation; its contribution is the taxonomy, the structured literature organization, and the synthesis of methods and challenges. The central claim is that this unified four-task view, together with coverage of emerging AI paradigms (ViTs, LLMs, diffusion models), provides a comprehensive and up-to-date foundation for camera-based VRU safety systems.

Significance. If taken as a reference survey, the paper is useful: the four-task organization is clear and sensible, the accompanying tables (Tables III–V) give a compact map of recent methods, and the challenge taxonomy in Section IV is practice-oriented. The review also explicitly covers less-common VRU classes (e-scooters, wheelchairs, mobility aids) in Section III-A, which is a genuine gap in many prior surveys. However, the significance is diminished by a mismatch between the advertised scope and the actual content. The abstract and conclusion promise systematic incorporation of large language models as one of three emerging AI paradigms, but no LLM-based method is reviewed anywhere in the body. Since the paper explicitly distinguishes itself from earlier surveys by filling this gap, the overstated claim affects the central contribution of comprehensiveness.

major comments (2)
  1. [Abstract and Section III (all subsections)] The abstract and the conclusion state that the survey 'systematically incorporates emerging AI paradigms, including Vision Transformers (ViTs), Large Language Models (LLMs), and diffusion models' and that these 'cutting-edge approaches' are a differentiator from prior surveys. Yet Section III contains no review, table entry, or cited method that uses an LLM for any of the four tasks. The only vision-language entries in Table III are Grounding DINO [44] and YOLO-World [45], which are open-vocabulary detectors, not LLM-based reasoning systems, and the text does not describe them as such. The conclusion repeats the claim ('ViTs, LLMs, and diffusion models'), but the body does not support it. This is load-bearing because the paper's stated novelty and the 'comprehensive' foundation it claims to provide depend in part on this coverage. The fix is either to add a substantive treatment of LLM-b
  2. [Section I, introductory positioning; Table I] The paper motivates itself by saying existing surveys 'primarily emphasize detection' and that key areas such as trajectory and intent prediction are 'often overlooked.' However, several cited reviews in Table I already cover trajectory prediction (e.g., [13]) and the broader VRU pipeline (e.g., [12]). The differentiation is weaker than claimed. The survey is still valuable as a unified treatment, but the contrast with prior work should be softened or made more precise; otherwise the contribution statement overstates the gap.
minor comments (5)
  1. [Title and header] The arXiv title is 'From Camera-Based Sensing to Reasoning: A Comprehensive Review Toward Proactive Vulnerable Road User Safety' but the full-text title is 'A Comprehensive Review on Artificial Intelligence Empowered Solutions for Enhancing Pedestrian and Cyclist Safety.' These should be aligned, especially since the final version may be indexed under one title.
  2. [References [128]-[131]] Entries such as 'Lorenzo et al.', 'Zhang et al.', and 'Sharma et al.' in the intent-recognition section are incomplete; they should be replaced with full author lists and stable bibliographic data.
  3. [Section III-C, GNN paragraph] The GNN paragraph says 'several Transformer-based models mentioned in the previous section such as HiVT [102] and AI-TP [95] integrate attention with graph structures.' AI-TP was earlier described under RNN-based models, not Transformer-based models, and the reference [95] is 'AI-TP: Attention-based interaction-aware trajectory prediction for autonomous driving.' The description is confusing and should be corrected.
  4. [Figure 1] The publication-trend figure is based on Google Scholar keyword searches but does not report the exact query, date of search, or filtering criteria. This makes the trend plot hard to reproduce and should be documented in the caption or text.
  5. [Section III-D, PCPA [119]] Reference [119] is cited as proposing 'a hybrid attention-based framework' for pedestrian action prediction, but the cited title indicates a benchmark paper ('Benchmark for evaluating pedestrian action prediction'). Please verify whether the attribution is to the benchmark or to a specific method, and cite accordingly.

Circularity Check

0 steps flagged

No circularity found: the survey's four-task taxonomy is a framing choice, not a derived result, and no prediction or parameter is fit from its own inputs.

full rationale

This is a literature survey, not a derivation with fitted parameters or benchmark predictions. The central contribution—organizing camera-based VRU safety into detection/classification, tracking/re-ID, trajectory prediction, and intent recognition/prediction—is a proposed taxonomy; the paper does not claim to prove this decomposition from first principles, nor does it run experiments to validate it. Statements such as 'these four tasks enable AI-based systems to sustain situational awareness' (Section II-B) are interpretive summaries of the surveyed literature, not reductions of outputs to inputs. No equations, fitted parameters, or benchmark numbers are used, so the main circularity patterns (self-definitional, fitted-input-called-prediction, ansatz-smuggling) do not apply. The paper cites several works by its own authors (e.g., [9], [52], [111]-[113]), but these are specific, peer-reviewed and benchmarked methods or background support, not a load-bearing 'uniqueness theorem' and not used to forbid alternative taxonomies or methods. Their inclusion is ordinary survey citation, not circularity. The most notable issue is internal consistency rather than circularity: the abstract says the survey 'systematically incorporates emerging AI paradigms, including ... LLMs,' but Section III's reviews of detection, tracking, trajectory prediction, and intent prediction contain no LLM-based method; only open-vocabulary detectors (Grounding DINO [44], YOLO-World [45]) appear. This undermines the advertised comprehensiveness and the introduction's framing that LLMs are 'largely unexamined in the current literature' and will be covered here, but an omitted body of work is an accuracy/completeness concern, not a derivation that equates the paper's conclusion to its own assumptions. Similarly, selecting papers that fit the four-task narrative is a survey-bias issue, not circularity. Because the paper makes no numerical prediction derived from its own inputs and its central claims are framing choices supported by external literature, the circularity score is 0.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

A survey has no free parameters or invented entities. Its load-bearing assumptions are the organizing taxonomy and the representativeness of the literature selection.

axioms (2)
  • domain assumption The proposed four-task pipeline (detection and classification, tracking and re-identification, trajectory prediction, intent recognition) is the correct decomposition for proactive VRU safety systems.
    Section III organizes the whole survey around this taxonomy without empirical or comparative evidence that it is the best structure.
  • domain assumption The selected cited papers are representative of the state of the art in each category.
    Tables III-V present representative methods, but no systematic search or inclusion criteria are described.

pith-pipeline@v1.3.0-alltime-deepseek · 127 in / 4005 out tokens · 57894 ms · 2026-08-04T13:24:39.766781+00:00 · methodology

0 comments
read the original abstract

Ensuring the safety of vulnerable road users (VRUs), such as pedestrians and cyclists, remains a critical challenge, as conventional infrastructure-based measures are often insufficient in dynamic urban environments. Recent advances in learning-based visual sensing systems have created new opportunities for more adaptive and context-aware VRU protection. This paper presents a comprehensive review of camera-based methods for proactive VRU safety, with a focus on developments over the past five years. Unlike prior surveys that primarily emphasize detection, we organize the literature into three interconnected components: visual perception (detection and classification), motion modeling (tracking and trajectory prediction), and behavior understanding (intent recognition and reasoning). These components form a unified hierarchical pipeline that enables early risk anticipation and timely intervention. Furthermore, this survey systematically incorporates emerging AI paradigms, including Vision Transformers (ViTs), Large Language Models (LLMs), and diffusion models, highlighting their roles in representation learning, uncertainty modeling, and semantic reasoning. Finally, we identify four key challenges specific to VRU safety, including data scarcity, behavioral uncertainty, edge deployment efficiency, and real-world sensing constraints, and discuss corresponding research directions. This work provides a unified foundation for the development of reliable, scalable, and deployable VRU safety systems.

Figures

Figures reproduced from arXiv: 2510.03314 by Bingzhang Wang, Chenxi Liu, Kehua Chen, Mehrdad Nasri, Muhammad Monjurul Karim, Shucheng Zhang, Yan Shi, Yinhai Wang, Yuang Zhang.

Figure 1
Figure 1. Figure 1: Publication Trends in AI Research for VRU and Pedestrian Safety [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Our Survey TABLE II DETAILED CLASSIFICATION OF VRUS AND ASSOCIATED SENSING CHALLENGES Class Subclasses Primary Sensing Challenges Pedestrian • Child • Adult • Elder • High variability in appearance and motion • Frequent occlusion in urban scenes • Different movement and reaction patterns Cyclist • Manual Bicycle • Motorized Bicycle • Fast and often non-linear motion • Occlusion by frames or hel… view at source ↗
Figure 3
Figure 3. Figure 3: Integrated Camera-Based Perception and Downstream Tasks for VRU Safety Enhancement [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Open Challenges and Future Directions for AI-Empowered VRU Sensing and Reasoning [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

148 extracted references · 16 linked inside Pith

  1. [1]

    Advancing traffic safety through the safe system approach: A systematic review,

    M. N. Khan and S. Das, “Advancing traffic safety through the safe system approach: A systematic review,” Accident Analysis & Prevention, vol. 199, p. 107518, 2024

    2024

  2. [2]

    Road traffic injuries,

    World Health Organization, “Road traffic injuries,” Dec. 2023, accessed on April 11, 2025. [Online]. Available: https://www.who.int/news-room/fact-sheets/ detail/road-traffic-injuries

    2023

  3. [3]

    U.s. department of transportation announces new guidance to improve safety for vulnerable road users under president’s bipartisan infrastructure law,

    U.S. Department of Transportation, “U.s. department of transportation announces new guidance to improve safety for vulnerable road users under president’s bipartisan infrastructure law,” Oct. 2022, accessed on April 11, 2025. [Online]. Available: https://www. transportation.gov/briefing-room

    2022

  4. [4]

    Nj named worst us state for pedestrian crashes — drugs and alcohol may be to blame,

    M. Sedacca, “Nj named worst us state for pedestrian crashes — drugs and alcohol may be to blame,” Mar. 2025, accessed on April 11, 2025. [Online]. Available: https://nypost.com/2025/03/08/ us-news/nj-most-dangerous-state-for-pedestrians-data/

    2025

  5. [5]

    Vulnerable road user safety assessment,

    Washington State Department of Transportation, “Vulnerable road user safety assessment,” Washington State Department of Transportation, Tech. Rep., 2023, accessed on April 11, 2025. [Online]. Available: https://targetzero.com/wp-content/uploads/ 2023/11/VRU-Safety-Assessment-2023.pdf

    2023

  6. [6]

    Cooperative collision avoidance at in- tersections: Algorithms and experiments,

    M. R. Hafner, D. Cunningham, L. Caminiti, and D. Del Vecchio, “Cooperative collision avoidance at in- tersections: Algorithms and experiments,”IEEE Trans- actions on Intelligent Transportation Systems, vol. 14, no. 3, pp. 1162–1175, 2013

    2013

  7. [7]

    Vulner- able road user detection for roadside-assisted safety protection: A comprehensive survey,

    Z. Zhang, C. Wei, G. Wu, and M. J. Barth, “Vulner- able road user detection for roadside-assisted safety protection: A comprehensive survey,”Applied Sciences, vol. 15, no. 7, p. 3797, 2025

    2025

  8. [8]

    Vehicle-to- everything (v2x) in the autonomous vehicles domain–a technical review of communication, sensor, and ai tech- nologies for road user safety,

    S. A. Yusuf, A. Khan, and R. Souissi, “Vehicle-to- everything (v2x) in the autonomous vehicles domain–a technical review of communication, sensor, and ai tech- nologies for road user safety,”Transportation Research Interdisciplinary Perspectives, vol. 23, p. 100980, 2024

    2024

  9. [9]

    Cooperative traffic signal assistance system for non- motorized users and disabilities empowered by com- puter vision and edge artificial intelligence,

    H. F. Yang, Y . Ling, C. Kopca, S. Ricord, and Y . Wang, “Cooperative traffic signal assistance system for non- motorized users and disabilities empowered by com- puter vision and edge artificial intelligence,”Trans- portation research part C: emerging technologies, vol. 145, p. 103896, 2022

    2022

  10. [10]

    Implementation and evaluation of a cooperative vehicle-to-pedestrian safety application,

    A. Tahmasbi-Sarvestani, H. N. Mahjoub, Y . P. Fallah, E. Moradi-Pari, and O. Abuchaar, “Implementation and evaluation of a cooperative vehicle-to-pedestrian safety application,”IEEE Intelligent Transportation Systems Magazine, vol. 9, no. 4, pp. 62–75, 2017

    2017

  11. [11]

    Roadside sensor systems for vulnerable road user protection: A review of methods and applications,

    T. Zhang, L. Cheng, T. Bang, L. Guo, M. Hajij, S. Cao, A. Harris, and M. Sartipi, “Roadside sensor systems for vulnerable road user protection: A review of methods and applications,”IEEE Access, 2025

    2025

  12. [12]

    Vulnerable road user detection and safety enhancement: A comprehensive survey,

    R. M. Silva, G. F. Azevedo, M. V . Berto, J. R. Rocha, E. C. Fidelis, M. V . Nogueira, P. H. Lisboa, and T. A. Almeida, “Vulnerable road user detection and safety enhancement: A comprehensive survey,”arXiv preprint arXiv:2405.19202, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Advancing vulnerable road users safety: Interdisci- plinary review on v2x communication and trajectory prediction,

    B. Abdi, S. Mirzaei, M. Adl, S. Hidajat, and A. Emadi, “Advancing vulnerable road users safety: Interdisci- plinary review on v2x communication and trajectory prediction,”IEEE Transactions on Intelligent Trans- portation Systems, 2024

    2024

  14. [14]

    Occlusion handling and multi-scale pedestrian detection based on deep learning: A review,

    F. Li, X. Li, Q. Liu, and Z. Li, “Occlusion handling and multi-scale pedestrian detection based on deep learning: A review,”IEEE Access, vol. 10, pp. 19 937–19 957, 2022

    2022

  15. [15]

    From handcrafted to deep features for pedestrian detection: A survey,

    J. Cao, Y . Pang, J. Xie, F. S. Khan, and L. Shao, “From handcrafted to deep features for pedestrian detection: A survey,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 9, pp. 4913–4934, 2021

    2021

  16. [16]

    Vulnerable road users,

    National Safety Council, “Vulnerable road users,” 2018, accessed on April 11, 2025. [Online]. Available: https://www.nsc.org/getattachment/ d5babee6-582d-4e66-804f-8d06f9b021a4/ t-vulnerable-road-users-147

    2018

  17. [17]

    Intersection safety challenge,

    U.S. Department of Transportation, “Intersection safety challenge,” 2024, accessed on April 11, 2025. [Online]. Available: https://its.dot.gov/isc/ JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 15

    2024

  18. [18]

    Detecting carried objects from sequences of walking pedestrians,

    D. Damen and D. Hogg, “Detecting carried objects from sequences of walking pedestrians,”IEEE transactions on pattern analysis and machine intelligence, vol. 34, no. 6, pp. 1056–1067, 2011

    2011

  19. [19]

    Histograms of oriented gra- dients for human detection,

    N. Dalal and B. Triggs, “Histograms of oriented gra- dients for human detection,” in2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05), vol. 1, 2005, pp. 886–893 vol. 1

    2005

  20. [20]

    Robust real-time face detec- tion,

    P. Viola and M. J. Jones, “Robust real-time face detec- tion,”International Journal of Computer Vision, vol. 57, no. 2, pp. 137–154, 5 2004

    2004

  21. [21]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” inComputer vision– ECCV 2014: 13th European conference, zurich, Switzer- land, September 6-12, 2014, proceedings, part v 13. Springer, 2014, pp. 740–755

    2014

  22. [22]

    Imagenet: A large-scale hierarchical im- age database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical im- age database,” in2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255

    2009

  23. [23]

    Ssd: Single shot multibox de- tector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg, “Ssd: Single shot multibox de- tector,” inLecture Notes in Computer Science. Springer International Publishing, 2016, pp. 21–37

    2016

  24. [24]

    Faster r-cnn: Towards real-time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp. 1137–1149, 2017

    2017

  25. [25]

    Efficientdet: Scalable and efficient object detection,

    M. Tan, R. Pang, and Q. V . Le, “Efficientdet: Scalable and efficient object detection,” 2020

    2020

  26. [26]

    Sparse r-cnn: End-to-end object detection with learnable pro- posals,

    P. Sun, R. Zhang, Y . Jiang, T. Kong, C. Xu, W. Zhan, M. Tomizuka, L. Li, Z. Yuan, C. Wanget al., “Sparse r-cnn: End-to-end object detection with learnable pro- posals,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 14 454–14 463

    2021

  27. [27]

    Pp-yoloe: An evolved version of yolo,

    S. Xu, X. Wang, W. Lv, Q. Chang, C. Cui, K. Deng, G. Wang, Q. Dang, S. Wei, Y . Duet al., “Pp-yoloe: An evolved version of yolo,” 2022

    2022

  28. [28]

    Convnext v2: Co-designing and scaling convnets with masked autoencoders,

    S. Woo, S. Debnath, R. Hu, X. Chen, Z. Liu, I. S. Kweon, and S. Xie, “Convnext v2: Co-designing and scaling convnets with masked autoencoders,” inPro- ceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, 2023, pp. 16 133–16 142

    2023

  29. [29]

    Internim- age: Exploring large-scale vision foundation models with deformable convolutions,

    W. Wang, J. Dai, Z. Chen, Z. Huang, Z. Li, X. Zhu, X. Hu, T. Lu, L. Lu, H. Liet al., “Internim- age: Exploring large-scale vision foundation models with deformable convolutions,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 14 408–14 419

    2023

  30. [30]

    Deformable detr: Deformable transformers for end-to- end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable detr: Deformable transformers for end-to- end object detection,”arXiv preprint arXiv:2010.04159, 2020

    Pith/arXiv arXiv 2010

  31. [31]

    Swin transformer v2: Scaling up capacity and resolution,

    Z. Liu, H. Hu, Y . Lin, Z. Yao, Z. Xie, Y . Wei, J. Ning, Y . Cao, Z. Zhang, L. Donget al., “Swin transformer v2: Scaling up capacity and resolution,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 12 009–12 019

    2022

  32. [32]

    Yolov11: An overview of the key architectural enhancements,

    R. Khanam and M. Hussain, “Yolov11: An overview of the key architectural enhancements,”arXiv preprint arXiv:2410.17725, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Yolov12: Attention- centric real-time object detectors,

    Y . Tian, Q. Ye, and D. Doermann, “Yolov12: Attention- centric real-time object detectors,”arXiv preprint arXiv:2502.12524, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Detrs beat yolos on real-time object detection,

    Y . Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y . Liu, and J. Chen, “Detrs beat yolos on real-time object detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 16 965–16 974

    2024

  35. [35]

    Fastervit: Fast vision transformers with hierarchical attention,

    A. Hatamizadeh, G. Heinrich, H. Yin, A. Tao, J. M. Alvarez, J. Kautz, and P. Molchanov, “Fastervit: Fast vision transformers with hierarchical attention,”arXiv preprint arXiv:2306.06189, 2023

    Pith/arXiv arXiv 2023

  36. [36]

    Edgevit: Efficient visual modeling for edge comput- ing,

    Z. Chen, F. Zhong, Q. Luo, X. Zhang, and Y . Zheng, “Edgevit: Efficient visual modeling for edge comput- ing,” inInternational conference on wireless algorithms, systems, and applications. Springer, 2022, pp. 393– 405

    2022

  37. [37]

    Maxvit: Multi-axis vision trans- former,

    Z. Tu, H. Talebi, H. Zhang, F. Yang, P. Milanfar, A. Bovik, and Y . Li, “Maxvit: Multi-axis vision trans- former,” inEuropean conference on computer vision. Springer, 2022, pp. 459–479

    2022

  38. [38]

    Diffusiondet: Diffusion model for object detection,

    S. Chen, P. Sun, Y . Song, and P. Luo, “Diffusiondet: Diffusion model for object detection,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 19 830–19 843

    2023

  39. [39]

    An improved yolov5 algorithm for vulnerable road user detection,

    W. Yang, X. Tang, K. Jiang, Y . Fu, and X. Zhang, “An improved yolov5 algorithm for vulnerable road user detection,”Sensors, vol. 23, no. 18, p. 7761, 2023

    2023

  40. [40]

    E-scooter rider detection and classification in dense urban environments,

    S. Gilroy, D. Mullins, E. Jones, A. Parsi, and M. Glavin, “E-scooter rider detection and classification in dense urban environments,”Results in Engineering, vol. 16, p. 100677, Dec. 2022

    2022

  41. [41]

    Deep learning architectures for skateboarder– pedestrian surrogate safety measures,

    C. E. Shourov, M. Sarkar, A. Jahangiri, and C. Paolini, “Deep learning architectures for skateboarder– pedestrian surrogate safety measures,”Future transportation, vol. 1, no. 2, pp. 387–413, 2021

    2021

  42. [42]

    Detection of e- scooter riders in naturalistic scenes,

    K. Apurv, R. Tian, and R. Sherony, “Detection of e- scooter riders in naturalistic scenes,”arXiv preprint arXiv:2111.14060, 2021

    Pith/arXiv arXiv 2021

  43. [43]

    Deep detection of people and their mobility aids for a hospital robot,

    A. Vasquez, M. Kollmitz, A. Eitel, and W. Burgard, “Deep detection of people and their mobility aids for a hospital robot,” in2017 European conference on mobile robots (ECMR). IEEE, 2017, pp. 1–7

    2017

  44. [44]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Suet al., “Grounding dino: Marrying dino with grounded pre-training for open-set object detection,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 38–55

    2024

  45. [45]

    Yolo-world: Real-time open-vocabulary ob- ject detection,

    T. Cheng, L. Song, Y . Ge, W. Liu, X. Wang, and Y . Shan, “Yolo-world: Real-time open-vocabulary ob- ject detection,” inProceedings of the IEEE/CVF Con- JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 16 ference on Computer Vision and Pattern Recognition, 2024, pp. 16 901–16 911

    2021

  46. [46]

    Pedestrian detection in low-light conditions: A comprehensive survey,

    B. Ghari, A. Tourani, A. Shahbahrami, and G. Gaydad- jiev, “Pedestrian detection in low-light conditions: A comprehensive survey,”Image and Vision Computing, p. 105106, 2024

    2024

  47. [47]

    Ms-detr: Multispectral pedes- trian detection transformer with loosely coupled fusion and modality-balanced optimization,

    Y . Xing, S. Yang, S. Wang, S. Zhang, G. Liang, X. Zhang, and Y . Zhang, “Ms-detr: Multispectral pedes- trian detection transformer with loosely coupled fusion and modality-balanced optimization,”IEEE Transac- tions on Intelligent Transportation Systems, 2024

    2024

  48. [48]

    Cross- modality fusion transformer for multispectral object detection,

    F. Qingyun, H. Dapeng, and W. Zhaokui, “Cross- modality fusion transformer for multispectral object detection,”arXiv preprint arXiv:2111.00273, 2021

    Pith/arXiv arXiv 2021

  49. [49]

    Guided attentive feature fusion for multispectral pedes- trian detection,

    H. Zhang, E. Fromont, S. Lef `evre, and B. Avignon, “Guided attentive feature fusion for multispectral pedes- trian detection,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, 2021, pp. 72–80

    2021

  50. [50]

    Multispectral fusion for object detection with cyclic fuse-and-refine blocks,

    H. Zhang, E. Fromont, S. Lefevre, and B. Avignon, “Multispectral fusion for object detection with cyclic fuse-and-refine blocks,” in2020 IEEE International conference on image processing (ICIP). IEEE, 2020, pp. 276–280

    2020

  51. [51]

    Improving multi- spectral pedestrian detection by addressing modality imbalance problems,

    K. Zhou, L. Chen, and X. Cao, “Improving multi- spectral pedestrian detection by addressing modality imbalance problems,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23– 28, 2020, Proceedings, Part XVIII 16. Springer, 2020, pp. 787–803

    2020

  52. [52]

    Illumination and temperature-aware multispectral networks for edge- computing-enabled pedestrian detection,

    Y . Zhuang, Z. Pu, J. Hu, and Y . Wang, “Illumination and temperature-aware multispectral networks for edge- computing-enabled pedestrian detection,”IEEE Trans- actions on Network Science and Engineering, vol. 9, no. 3, pp. 1282–1295, 2021

    2021

  53. [53]

    All-weather pedestrian detection based on double-stream multispec- tral network,

    C.-H. Hsia, H.-C. Peng, and H.-T. Chan, “All-weather pedestrian detection based on double-stream multispec- tral network,”Electronics, vol. 12, no. 10, p. 2312, 2023

    2023

  54. [54]

    Illumination- aware faster r-cnn for robust multispectral pedestrian detection,

    C. Li, D. Song, R. Tong, and M. Tang, “Illumination- aware faster r-cnn for robust multispectral pedestrian detection,”Pattern Recognition, vol. 85, pp. 161–171, 2019

    2019

  55. [55]

    Pedestrian detection based on light perception fusion of visible and thermal im- ages,

    G. Li, W. Lai, and X. Qu, “Pedestrian detection based on light perception fusion of visible and thermal im- ages,”Optics & Laser Technology, vol. 156, p. 108466, 2022

    2022

  56. [56]

    Mask-guided attention network for occluded pedestrian detection,

    Y . Pang, J. Xie, M. H. Khan, R. M. Anwer, F. S. Khan, and L. Shao, “Mask-guided attention network for occluded pedestrian detection,” 2019

    2019

  57. [57]

    Nms by repre- sentative region: Towards crowded pedestrian detection by proposal pairing,

    X. Huang, Z. Ge, Z. Jie, and O. Yoshie, “Nms by repre- sentative region: Towards crowded pedestrian detection by proposal pairing,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 10 750–10 759

    2020

  58. [58]

    Imagine the unseen: Occluded pedestrian detection via adversarial feature completion,

    S. Zhang, M. Ji, Y . Li, and J. Yang, “Imagine the unseen: Occluded pedestrian detection via adversarial feature completion,”arXiv preprint arXiv:2405.01311, 2024

    Pith/arXiv arXiv 2024

  59. [59]

    Mapd: An improved multi-attribute pedestrian detection in a crowd,

    Y . Wang, C. Han, G. Yao, and W. Zhou, “Mapd: An improved multi-attribute pedestrian detection in a crowd,”Neurocomputing, vol. 432, pp. 101–110, 2021

    2021

  60. [60]

    Beta r-cnn: Look- ing into pedestrian detection from another perspective,

    Z. Xu, B. Li, Y . Yuan, and A. Dang, “Beta r-cnn: Look- ing into pedestrian detection from another perspective,” Advances in Neural Information Processing Systems, vol. 33, pp. 19 953–19 963, 2020

    2020

  61. [61]

    Detection in crowded scenes: One proposal, multiple predictions,

    X. Chu, A. Zheng, X. Zhang, and J. Sun, “Detection in crowded scenes: One proposal, multiple predictions,” 2020

    2020

  62. [62]

    Pedhunter: Occlusion robust pedestrian detector in crowded scenes,

    C. Chi, S. Zhang, J. Xing, Z. Lei, S. Z. Li, and X. Zou, “Pedhunter: Occlusion robust pedestrian detector in crowded scenes,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 07, 2020, pp. 10 639–10 646

    2020

  63. [63]

    Adaptive pattern-parameter matching for robust pedestrian de- tection,

    M. Liu, C. Zhu, J. Wang, and X.-C. Yin, “Adaptive pattern-parameter matching for robust pedestrian de- tection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 3, 2021, pp. 2154– 2162

    2021

  64. [64]

    Oaf-net: An occlusion-aware anchor-free network for pedestrian detection in a crowd,

    Q. Li, Y . Su, Y . Gao, F. Xie, and J. Li, “Oaf-net: An occlusion-aware anchor-free network for pedestrian detection in a crowd,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 11, pp. 21 291– 21 300, 2022

    2022

  65. [65]

    Autopedestrian: An automatic data aug- mentation and loss function search scheme for pedes- trian detection,

    Y . Tang, B. Li, M. Liu, B. Chen, Y . Wang, and W. Ouyang, “Autopedestrian: An automatic data aug- mentation and loss function search scheme for pedes- trian detection,”IEEE transactions on image process- ing, vol. 30, pp. 8483–8496, 2021

    2021

  66. [66]

    Tracking pedestrian heads in dense crowd,

    R. Sundararaman, C. De Almeida Braga, E. Marchand, and J. Pettre, “Tracking pedestrian heads in dense crowd,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 3865–3875

    2021

  67. [67]

    Detr for crowd pedestrian detection,

    M. Lin, C. Li, X. Bu, M. Sun, C. Lin, J. Yan, W. Ouyang, and Z. Deng, “Detr for crowd pedestrian detection,” 2021

    2021

  68. [68]

    Optimal proposal learning for deployable end-to-end pedestrian detection,

    X. Song, B. Chen, P. Li, J.-Y . He, B. Wang, Y . Geng, X. Xie, and H. Zhang, “Optimal proposal learning for deployable end-to-end pedestrian detection,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 3250–3260

    2023

  69. [69]

    Uncertainty-guided cross-modal learning for robust multispectral pedestrian detection,

    J. U. Kim, S. Park, and Y . M. Ro, “Uncertainty-guided cross-modal learning for robust multispectral pedestrian detection,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 3, pp. 1510–1523, 2021

    2021

  70. [70]

    Causal mode multiplexer: A novel framework for unbi- ased multispectral pedestrian detection,

    T. Kim, S. Shin, Y . Yu, H. G. Kim, and Y . M. Ro, “Causal mode multiplexer: A novel framework for unbi- ased multispectral pedestrian detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 26 784–26 793

    2024

  71. [71]

    Simple online and realtime tracking,

    A. Bewley, Z. Ge, L. Ott, F. Ramos, and B. Upcroft, “Simple online and realtime tracking,” in2016 IEEE International Conference on Image Processing (ICIP). IEEE, 2016, pp. 3464–3468. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 17

    2016

  72. [72]

    Simple online and realtime tracking with a deep association metric,

    N. Wojke, A. Bewley, and D. Paulus, “Simple online and realtime tracking with a deep association metric,” in2017 IEEE International Conference on Image Pro- cessing (ICIP). IEEE, 2017, pp. 3645–3649

    2017

  73. [73]

    Tracking ob- jects as points,

    X. Zhou, D. Wang, and P. Kr ¨ahenb¨uhl, “Tracking ob- jects as points,” inEuropean Conference on Computer Vision (ECCV). Springer, 2020, pp. 474–490

    2020

  74. [74]

    Fairmot: On the fairness of detection and re- identification in multiple object tracking,

    Y . Zhang, C. Wang, X. Wang, W. Zeng, and W. Liu, “Fairmot: On the fairness of detection and re- identification in multiple object tracking,”International Journal of Computer Vision, vol. 129, no. 11, pp. 3069– 3087, 2021

    2021

  75. [75]

    Towards real-time multi-object tracking,

    Z. Wang, L. Zheng, Y . Liu, Y . Li, and S. Wang, “Towards real-time multi-object tracking,” inEuropean Conference on Computer Vision (ECCV). Springer, 2020, pp. 107–122

    2020

  76. [76]

    Quasi- dense similarity learning for multiple object tracking,

    J. Pang, K. Li, Y . Zhang, Y . Li, and C. Lu, “Quasi- dense similarity learning for multiple object tracking,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  77. [77]

    Rethinking the competition between detection and reid in multi-object tracking,

    C. Liang, Z. Zhang, X. Zhou, B. Li, S. Zhu, and W. Hu, “Rethinking the competition between detection and reid in multi-object tracking,” 2022

    2022

  78. [78]

    Bytetrack: Multi-object tracking by associating every detection box,

    Y . Zhang, P. Sun, Y . Jiang, D. Yu, C. Weng, Z. Yuan, and P. Luo, “Bytetrack: Multi-object tracking by associating every detection box,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 9839–9848

    2021

  79. [79]

    Observation-centric sort: Rethinking sort for robust multi-object tracking,

    J. Cao, J. Pang, X. Weng, R. Khirodkar, and K. Kitani, “Observation-centric sort: Rethinking sort for robust multi-object tracking,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2023, pp. 973–983

    2023

  80. [80]

    Bot- sort: Robust associations multi-pedestrian tracking,

    N. Aharoni, B.-Z. Bobrovsky, and Y . Ben-Shabat, “Bot- sort: Robust associations multi-pedestrian tracking,” arXiv preprint arXiv:2206.14651, 2022

    Pith/arXiv arXiv 2022

Showing first 80 references.