Pith. sign in

REVIEW 3 major objections 5 minor 51 references

Thermal Detection of People with Mobility Restrictions for Barrier Reduction at Traffic Lights Controlled Intersections

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a thermal-camera detector can identify pedestrians with mobility restrictions at intersections and adapt traffic lights, extending green time and triggering audio cues for those who need them.

desk verdict A genuinely useful thermal dataset for mobility-restricted pedestrians, wrapped in a modest detector update and an under-supported controller simulation. read the letter →

arxiv 2505.08568 v2 pith:LMH7H5QS submitted 2025-05-13 cs.CV

Xiao Ni , Carsten Kuehnel , Xiaoyi Jiang This is my paper · ORCID
classification cs.CV
keywords barrier-freeintersectionthermalimagesobjectdetectiondatasetpedestrianmobilityrestrictionsadaptivetrafficlightcontrolYOLO
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

This paper tries to establish that traffic-light intersections can become barrier-free without requiring pedestrians to carry any device: thermal cameras watch the crosswalk, a detector identifies people who need extra time, and the light controller extends green phases and switches on audible guidance accordingly. To support that, the authors built a new annotated thermal dataset of 11,196 images with twelve classes spanning distinct mobility aids and burdens, captured across seasons and lighting conditions at 640x512 resolution. They then designed YOLO-Thermal, a YOLOv8-based detector that adds Triplet-Attention, SPD-Conv, SPPFCSPC, and Quality Focal Loss to cope with the low resolution and missing texture of thermal imagery. On this dataset the model reports 89.1% average precision and 90.1 FPS, and the controller strategy with two-frame exit validation reports a 95.4% real success rate. The sympathetic reading is that thermal-only sensing can deliver accessibility benefits that RGB systems cannot, while sidestepping the privacy concerns that block camera-based traffic systems.

What carries the argument

The load-bearing pieces are the dataset, the detector, and the controller. TD4PWMR supplies the annotated thermal images and the twelve semantic classes that define what the system must see. YOLO-Thermal is the detector: it starts from YOLOv8 and adds Triplet-Attention, a three-branch attention module that captures cross-dimension dependencies; SPD-Conv, a space-to-depth convolution that preserves fine detail when downsampling; SPPFCSPC, a spatial-pyramid-pooling block with cross-stage partial connections that captures multi-scale context; and Quality Focal Loss, a classification loss with continuous IoU targets that helps with class imbalance. The controller groups the twelve classes into three pedestrian categories, applies fixed per-category green-time extensions, and uses a multi-frame validation rule that only ends an extension after the pedestrian is absent for N consecutive processed frames. This validation mechanism is what converts sporadic detection failures into a reliably long green phase, at the cost of latency that the paper measures and tunes through N.

What would settle it

Compute per-class average precision for the five mobility-aid classes on the held-out test portion of TD4PWMR; if wheelchair, rollator, crutches, blind stick, or walking stick are frequently confused with one another, the adaptive controller cannot reliably give the correct green-time and audio response to the group it is trying to help.

Watch

Extended reading notes

Core claim

The paper's central claim is that a thermal-only object detector can recognize pedestrians with mobility restrictions well enough to drive adaptive traffic-light control. The authors introduce TD4PWMR, a dataset of 11,196 manually annotated thermal frames showing people with wheelchairs, rollators, crutches, blind sticks, walking sticks, luggage, strollers, bicycles, trolleys, and dogs, plus ordinary pedestrians and cars. They report that YOLO-Thermal reaches an AP of 89.1%, AP@50 of 95.1%, and 90.1 FPS on this dataset, beating the YOLO and RT-DETR models they retrained for comparison. The traffic-light controller assigns the classes to three groups — walking impairments, visual impairments, and mobility burden — and extends green time by up to 6, 8, and 3 seconds respectively, with audible guidance for the visual group; when the detector temporarily loses a person, the green extension continues until the person stays absent for N consecutive frames, and N=2 is reported to give a 95.4% real success rate at 1.2 seconds of added latency. The paper concludes that this combination significantly improves intersection accessibility while keeping traffic flow efficient.

Load-bearing premise

The load-bearing premise is that a thermal picture taken from a pole at the crossing reliably shows which mobility aid a person uses, because the paper reports only an overall accuracy number and never shows how well the detector does on each type of aid separately.

Editorial extensions

If this is right

  • At a four-way intersection with eight thermal cameras, the detector can run on edge hardware while the controller checks whether a detected pedestrian's position falls inside the pre-defined crosswalk polygon.
  • A person classified as walking-impaired adds up to 6 seconds of green time, a visually impaired person up to 8 seconds plus an audible signal, and a person with mobility burden up to 3 seconds; if several are present, visual impairment has priority, then walking impairment, then mobility burden.
  • Using N=2 consecutive frames without a detection before declaring the crosswalk empty yields a reported 95.4% real success rate with 1.2 seconds of added latency, meeting the 95% design target while limiting the impact on traffic.
  • Because the sensor is thermal, the system does not depend on visible-light conditions and does not record identifiable facial detail, so it can operate at night, in poor weather, and in privacy-sensitive urban spaces.
  • The LayerCAM visualizations in the paper support the annotation choice of bounding the person and their aid as one object, indicating the detector learns the spatial relation between the two.

Reading between the lines

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

  • The paper reports only aggregate average precision, so an unstated next step is to publish per-class AP for the five mobility-aid classes; a confusion between blind stick and walking stick would produce the wrong controller response, sending an audio cue to someone who does not need it or withholding it from someone who does.
  • The dataset appears to have been captured at one instrumented intersection, so the claimed behavior may depend on that site's camera angles and thermal environment; a cross-site evaluation would show how much retraining or domain adaptation the system needs.
  • A field test could measure actual crossing durations for each user group and compare them with the paper's 95th-percentile estimates, since those estimates set the 6-, 8-, and 3-second extension caps that the controller relies on.
  • Whether mobility-burden classes such as luggage, dog, or bicycle should trigger extensions is as much a policy decision as a detection one; the three-second cap suggests the authors intend only a minor accommodation, and a deployment could tune or disable those classes to protect traffic flow.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a thermal-imaging-based adaptive traffic light system for pedestrians with mobility restrictions. It introduces TD4PWMR, a 11,196-image thermal dataset with 12 manually annotated classes, and YOLO-Thermal, a YOLOv8 variant augmented with Triplet-Attention, SPD-Conv, SPPFCSPC, and Quality Focal Loss. The authors report that YOLO-Thermal achieves the highest aggregate COCO AP on their dataset and that a multi-frame validation strategy, with parameter N selected via ablation, yields a 'real success rate' above 95%. The controller extends green time by 6, 8, or 3 seconds depending on whether the detected person has a walking impairment, visual impairment, or mobility burden, and triggers an auditory signal for visually impaired pedestrians.

Significance. If the claims are supported, the main contribution is a useful and privacy-preserving thermal dataset for an underrepresented application, together with a detector whose aggregate accuracy is competitive. The public dataset and code are concrete assets for the community. However, the paper's central accessibility claim is not established by the current evidence: the controller's actions are class-specific, yet only aggregate detection AP is reported, and the controller's effectiveness rests on an underspecified simulation that is tuned to the authors' own 95% target. The work is therefore potentially significant but requires substantially stronger evidence before the stated conclusions are justified.

major comments (3)
  1. [Section VI-A, Table III and Section III (Fig. 2)] The central accessibility claim depends on correct per-class identification because Table II maps each class to a different controller action (6 s vs 8 s vs 3 s and audio). However, Table III reports only aggregate COCO AP, and Fig. 2 shows heavy class imbalance between the majority class and the mobility-restricted classes. The claimed superiority is also marginal: the improvement over YOLOv9-E is about 0.3 AP (89.1 vs 88.8), with no error bars, multiple seeds, or significance testing. Without per-class AP or a confusion matrix, a model can achieve high aggregate AP while confusing, for example, a blindstick user with a walking-stick user, which would trigger the wrong extension time and omit the required auditory cue for the target population. Please report per-class AP, a confusion matrix, and a discussion of class-specific error modes; this evidence is load-bearing for the paper's stated application.
  2. [Section VI-B, Table V and Section IV] The selection of N=2 and the resulting 'real success rate' of 95.4% come from a simulation whose constitutive assumptions are not specified. The text defines success rate and real success rate but does not describe the pedestrian crossing-time distribution, the detector miss-rate model, the pedestrian arrival process, or the traffic-light timing model used to compute Table V. Choosing N to satisfy the authors' 95% design criterion within this unspecified simulation is circular: the simulation could be the source of the conclusion, not a validation of it. The authors must provide the full simulation specification, justify its parameters independently of the target metric, and include sensitivity analyses over crossing speeds, detection failure rates, and signal timing before the controller's effectiveness claim is supported.
  3. [Section IV] The green-time extensions of 6 s, 8 s, and 3 s are derived from 'our measurements' that are not reported anywhere in the manuscript. No measurement protocol, sample size, crossing-time distributions, or analysis underlying the 5.8 s, 7.9 s, and 2.7 s differences are given. Because these three values are exactly the controller outputs evaluated in Section VI-B, this missing evidence is directly load-bearing for the accessibility claim. Please add a reproducible description of the crossing-time measurements and the statistical basis for the 95th-percentile values, or clearly re-position these numbers as design assumptions.
minor comments (5)
  1. [Section V, first paragraph] The sentence 'we integrate integrate the SPPFCSPC module' contains a duplicated verb and should be corrected.
  2. [Table II] The table cell boundaries appear corrupted in the current version, with several class rows merged into single entries; please regenerate the table so that each class and its adaptation strategy are legible.
  3. [Section VI-A] The discussion of APval_S and APval_L is garbled: the text says 'APval_L and APval_L split the evaluation' and does not clearly state which column corresponds to small versus large objects. Please fix the notation and add explicit column labels.
  4. [Section V-D] In the description of Quality Focal Loss, 'p donates the predicted confidence score' should read 'denotes'.
  5. [Section VI-B] The phrase 'addtional time' should read 'additional time'; this typo appears in the definition of the success rate.

Circularity Check

2 steps flagged · score 5.0 of 10

The controller's 95% success claim is tuned to its own design criterion, and the LayerCAM annotation validation is self-referential; the detector comparison itself is not circular.

  1. fitted input called prediction [Section VI-B (Ablation Study, Table V); multi-frame parameter N introduced in Section IV]
    "At N = 2, performance improves significantly. The success rate rises to 77.2%, and the real success rate reaches 95.4%, effectively satisfying the design criterion of supporting at least 95% of people with mobility restrictions."

    N is the free parameter of the multi-frame validation scheme, and the paper says its optimization is 'crucial' before reporting Table V. Choosing N=2 because it is the first value whose simulated real success rate crosses the 95% design criterion means the reported 'effectively satisfying' result is the optimization target, not an independent outcome. The same simulation is used both to select N and to validate the controller, and the 6/8/3 s extension inputs come from the paper's own unshown crossing-time measurements, so the accessibility success metric is partly forced by the parameters it is supposed to demonstrate.

  2. other [Section VI-C (Visualization)]
    "This observation supports our hypothesis that the relative position between the person and the associated object is critical for accurate detection. Moreover, the consistent activation patterns validate our annotation strategy, which treats the individual and the associated object as a single instance."

    The model was trained with ground-truth boxes that by definition enclose the person together with the mobility aid as one instance. The LayerCAM heatmaps therefore reflect the supervision imposed by that annotation strategy, so using them to 'validate' the same strategy is self-referential: the evidence is produced by the very labeling choice it is said to confirm. An independent test would require comparing models trained under alternative annotation schemes or evaluating per-class accuracy on the minority mobility-aid classes.

full rationale

The object-detection comparison is not circular: YOLO-Thermal and the SOTA baselines are all trained on the TD4PWMR training split and evaluated on the held-out test split at the same 640x640 input size, with COCO-style AP reported on that test set. No self-citation is load-bearing (reference [3] is an unrelated prior paper by a co-author). The circularity is localized in the controller evaluation. The 6/8/3 s green extensions are justified only by 'Our measurements indicate...' with no data or method shown, and the simulated 'real success rate' is defined against those same extension amounts, so the 95% success claim is partly built from its own inputs. More concretely, N is selected from the ablation table precisely because N=2 yields a real success rate above the 95% design criterion; reporting that same table as evidence that the system 'effectively satisfies' the criterion is fitting the parameter to the evaluation metric rather than independently predicting it. The LayerCAM 'validation' of the combined person-plus-aid annotation is also self-referential for the reasons in the second step. Missing per-class AP and the unshown crossing-time measurements are correctness/support gaps rather than circularity, but they compound the controller-claim problem.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central system depends on empirically fitted timing values and a tuned validation parameter, plus a set of domain assumptions about thermal discriminability and simulation validity. No new physical entities are introduced.

free parameters (4)
  • Green time extension for walking impairments = 6 s
    Derived from a claimed 95th-percentile crossing-time measurement of 5.8 s, rounded to integer seconds; no measurement protocol is provided (Section IV).
  • Green time extension for visual impairments = 8 s
    Derived from a claimed 7.9 s additional crossing time for 95% of visually impaired pedestrians, rounded; no measurement protocol (Section IV).
  • Green time extension for mobility burden = 3 s
    Derived from a claimed 2.7 s additional crossing time for 95% of pedestrians with mobility burden, rounded; no measurement protocol (Section IV).
  • Multi-frame validation count N = 2
    Selected from an ablation to make the simulated real success rate reach 95%, the authors' own success criterion (Table V, Section VI-B).
assumptions (3)
  • domain assumption The 95% design principle is an appropriate basis for setting traffic-light extensions.
    Invoked in Section IV with citation [41]; no justification that the 95th percentile crossing time is the right safety target.
  • domain assumption The simulated success rate, computed from single-frame detections and the multi-frame validation rule, is a valid proxy for real-world intersection safety.
    The simulation in Section VI-B assumes a relationship between per-frame detection success and safe crossing outcomes, without testing in real traffic.
  • domain assumption Thermal images at 640x512 provide enough information to distinguish the twelve annotated classes, including individual mobility aids.
    The paper relies on this in Section III and Section VI, but reports only aggregate AP and no per-class results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Thermal Detection of People with Mobility Restrictions for Barrier Reduction at Traffic Lights Controlled Intersections." pith.science (2026). https://pith.science/paper/LMH7H5QS

@misc{pith2026250508568,
  author       = {Pith},
  title        = {Pith review of: Thermal Detection of People with Mobility Restrictions for Barrier Reduction at Traffic Lights Controlled Intersections},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LMH7H5QS}},
  note         = {Machine review of arXiv:2505.08568}
}
read the original abstract

Rapid advances in deep learning for computer vision have driven the adoption of RGB camera-based adaptive traffic light systems to improve traffic safety and pedestrian comfort. However, these systems often overlook the needs of people with mobility restrictions. Moreover, the use of RGB cameras presents significant challenges, including limited detection performance under adverse weather or low-visibility conditions, as well as heightened privacy concerns. To address these issues, we propose a fully automated, thermal detector-based traffic light system that dynamically adjusts signal durations for individuals with walking impairments or mobility burden and triggers the auditory signal for visually impaired individuals, thereby advancing towards barrier-free intersection for all users. To this end, we build the thermal dataset for people with mobility restrictions (TD4PWMR), designed to capture diverse pedestrian scenarios, particularly focusing on individuals with mobility aids or mobility burden under varying environmental conditions, such as different lighting, weather, and crowded urban settings. While thermal imaging offers advantages in terms of privacy and robustness to adverse conditions, it also introduces inherent hurdles for object detection due to its lack of color and fine texture details and generally lower resolution of thermal images. To overcome these limitations, we develop YOLO-Thermal, a novel variant of the YOLO architecture that integrates advanced feature extraction and attention mechanisms for enhanced detection accuracy and robustness in thermal imaging. Experiments demonstrate that the proposed thermal detector outperforms existing detectors, while the proposed traffic light system effectively enhances barrier-free intersection. The source codes and dataset are available at https://github.com/leon2014dresden/YOLO-THERMAL.

Figures

Figures reproduced from arXiv: 2505.08568 by the authors.

Figure 1
Figure 1. One representative example from each class in the proposed TD4PWMR dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Class distribution, highlighting the significant imbalance in the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of raw images (top) and corresponding LayerCAM heatmaps (bottom). Warmer colors indicate regions of higher activation. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 43 canonical work pages

  1. [1]

    Transportation 5.0: The dao to safe, secure, and sustainable intelligent transportation systems,

    F.-Y . Wang, Y . Lin, P. A. Ioannou, L. Vlacic, X. Liu, A. Eskandarian, Y . Lv, X. Na, D. Cebon, J. Ma, L. Li, and C. Olaverri-Monreal, “Transportation 5.0: The dao to safe, secure, and sustainable intelligent transportation systems,” IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 10, pp. 10 262–10 278, 2023

    work page 2023

  2. [2]

    Real-time IoT-based connected vehicle infrastructure for intelligent transportation safety,

    N. Sharma and R. D. Garg, “Real-time IoT-based connected vehicle infrastructure for intelligent transportation safety,” IEEE Transactions on Intelligent Transportation Systems , vol. 24, no. 8, pp. 8339–8347, 2023

    work page 2023

  3. [3]

    Recognition of traffic lights in live video streams on mobile devices,

    J. Roters, X. Jiang, and K. Rothaus, “Recognition of traffic lights in live video streams on mobile devices,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 21, no. 10, pp. 1497–1511, 2011

    work page 2011

  4. [4]

    Barriers to accessibility of urban roads by persons with disabilities: A review of the literature,

    G. Asiedu-Ampem, A. Danso, J. Ayarkwa, D. Obeng-Atuah, E. Tudzi, and A. Afful, “Barriers to accessibility of urban roads by persons with disabilities: A review of the literature,” Journal of Transport & Health , vol. 39, p. 101935, 2024

    work page 2024

  5. [5]

    Assistive intelligent transportation systems: The need for user localization and anonymous disability identification,

    D. Fernandez-Llorca, R. Quintero Minguez, I. Parra Alonso, C. Fernan- dez Lopez, I. Garcia Daza, M. A. Sotelo, and C. A. Cordero, “Assistive intelligent transportation systems: The need for user localization and anonymous disability identification,” IEEE Intelligent Transportation Systems Magazine, vol. 9, no. 2, pp. 25–40, 2017

    work page 2017

  6. [6]

    A novel signal- scheduling algorithm with quality-of-service provisioning for an isolated intersection,

    R. Wunderlich, C. Liu, I. Elhanany, and T. Urbanik, “A novel signal- scheduling algorithm with quality-of-service provisioning for an isolated intersection,” IEEE Transactions on Intelligent Transportation Systems , vol. 9, no. 3, pp. 536–547, 2008

    work page 2008

  7. [7]

    Efficient and safe strategies for intersection management: A review,

    J. Wang, X. Guo, and X. Yang, “Efficient and safe strategies for intersection management: A review,” Sensors, vol. 21, no. 9, 2021

    work page 2021

  8. [8]

    Disparities in road crash mortality among pedestrians using wheelchairs in the usa: results of a capture–recapture analysis,

    J. D. Kraemer and C. S. Benton, “Disparities in road crash mortality among pedestrians using wheelchairs in the usa: results of a capture–recapture analysis,” BMJ Open, vol. 5, no. 11, 2015. [Online]. Available: https://bmjopen.bmj.com/content/5/11/e008396

    work page 2015

Show all 51 references
  1. [9]

    A real-time object detecting and tracking system for outdoor night surveillance,

    K. Huang, L. Wang, T. Tan, and S. Maybank, “A real-time object detecting and tracking system for outdoor night surveillance,” Pattern Recognition, vol. 41, no. 1, pp. 432–444, 2008

    2008

  2. [10]

    Image-adaptive YOLO for object detection in adverse weather conditions,

    W. Liu, G. Ren, R. Yu, S. Guo, J. Zhu, and L. Zhang, “Image-adaptive YOLO for object detection in adverse weather conditions,” in AAAI Conference on Artificial Intelligence , 2021

    2021

  3. [11]

    Prior-based domain adaptive object detection for hazy and rainy conditions,

    V . A. Sindagi, P. Oza, R. Yasarla, and V . M. Patel, “Prior-based domain adaptive object detection for hazy and rainy conditions,” in European Conference on Computer Vision , 2020, pp. 763–780

    2020

  4. [12]

    A comprehensive review of object detection with deep learning,

    R. Kaur and S. Singh, “A comprehensive review of object detection with deep learning,” Digital Signal Processing , vol. 132, p. 103812, 2023

    2023

  5. [13]

    Thermal cameras and applications: A survey,

    R. Gade and T. Moeslund, “Thermal cameras and applications: A survey,” Machine Vision and Applications , vol. 25, pp. 245–262, 01 2014

    2014

  6. [14]

    A systematic literature review on object detection using near infrared and thermal images,

    N. Bustos, M. Mashhadi, S. K. Lai-Yuen, S. Sarkar, and T. K. Das, “A systematic literature review on object detection using near infrared and thermal images,” Neurocomputing, vol. 560, p. 126804, 2023

    2023

  7. [15]

    YOLOv6 v3.0: A full-scale reloading,

    C. Li, L. Li, Y . Geng, H. Jiang, M. Cheng, B. Zhang, Z. Ke, X. Xu, and X. Chu, “YOLOv6 v3.0: A full-scale reloading,” ArXiv, vol. abs/2301.05586, 2023

    2023 arXiv

  8. [16]

    No more strided convolutions or pooling: A new CNN building block for low-resolution images and small objects,

    R. Sunkara and T. Luo, “No more strided convolutions or pooling: A new CNN building block for low-resolution images and small objects,” in Machine Learning and Knowledge Discovery in Databases. Springer Nature Switzerland, 2023, pp. 443–459

    2023

  9. [17]

    Rotate to attend: Convolutional triplet attention module,

    D. Misra, T. Nalamada, A. U. Arasanipalai, and Q. Hou, “Rotate to attend: Convolutional triplet attention module,” in IEEE Winter Conference on Applications of Computer Vision , 2021, pp. 3138–3147

    2021

  10. [18]

    Background-subtraction using contour- based fusion of thermal and visible imagery,

    J. W. Davis and V . Sharma, “Background-subtraction using contour- based fusion of thermal and visible imagery,” Computer Vision and Image Understanding, vol. 106, no. 2, pp. 162–182, 2007

    2007

  11. [19]

    Accessed: 2025-02-06

    Flir thermal dataset. Accessed: 2025-02-06. [Online]. Available: https://www.flir.com/oem/adas/adas-dataset-form/

    2025

  12. [20]

    Kaist multi-spectral day/night data set for autonomous and as- sisted driving,

    Y . Choi, N. Kim, S. Hwang, K. Park, J. S. Yoon, K. An, and I. S. Kweon, “Kaist multi-spectral day/night data set for autonomous and as- sisted driving,”IEEE Transactions on Intelligent Transportation Systems, vol. 19, no. 3, pp. 934–948, 2018

    2018

  13. [21]

    Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather,

    M. Bijelic, T. Gruber, F. Mannan, F. Kraus, W. Ritter, K. Dietmayer, and F. Heide, “Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 11 679–11 689

    2020

  14. [22]

    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,” in Proc. of the IEEE Eur. Conf. on Mobile Robotics , 2017

    2017

  15. [23]

    Deep 3D perception of people and their mobility aids,

    M. Kollmitz, A. Eitel, A. Vasquez, and W. Burgard, “Deep 3D perception of people and their mobility aids,” Robotics and Autonomous Systems , vol. 114, pp. 29–40, 2019

    2019

  16. [24]

    Mobility aids detection using convolution neural network (CNN),

    A. Mukhtar, M. J. Cree, J. B. Scott, and L. Streeter, “Mobility aids detection using convolution neural network (CNN),” in International Conference on Image and Vision Computing New Zealand , 2018, pp. 1–5

    2018

  17. [25]

    Improving pedestrian’s cross- walk accessibility through digital fencing,

    F. Bustos, J. Gonc ¸alves, and J. P. Coelho, “Improving pedestrian’s cross- walk accessibility through digital fencing,” in Symposium of Applied Science for Young Researchers , 2022

    2022

  18. [26]

    Mobile devices sup- porting people with special needs,

    T. Stefanov, S. Varbanova, and M. Stefanova, “Mobile devices sup- porting people with special needs,” International Journal of Advanced Computer Science and Applications , vol. 13, 2022

    2022

  19. [27]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in IEEE Conference on Computer Vision and Pattern Recognition , 2016, pp. 779–788

    2016

  20. [28]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 42, no. 2, pp. 318–327, 2020

    2020

  21. [29]

    SSD: Single shot multibox detector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg, “SSD: Single shot multibox detector,” in European Conference on Computer Vision, 2016, pp. 21–37

    2016

  22. [30]

    DETRs beat YOLOs on real-time object detection,

    W. Lv, S. Xu, Y . Zhao, G. Wang, J. Wei, C. Cui, Y . Du, Q. Dang, and Y . Liu, “DETRs beat YOLOs on real-time object detection,” IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 16 965– 16 974, 2023

    2023

  23. [31]

    YOLOv4: Op- timal speed and accuracy of object detection,

    A. Bochkovskiy, C.-Y . Wang, and H.-Y . M. Liao, “YOLOv4: Op- timal speed and accuracy of object detection,” arXiv preprint arXiv:2004.10934, 2020

    2004 arXiv

  24. [32]

    Scaled-YOLOv4: Scaling cross stage partial network,

    C.-Y . Wang, A. Bochkovskiy, and H.-Y . M. Liao, “Scaled-YOLOv4: Scaling cross stage partial network,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 13 024–13 033

    2021

  25. [33]

    YOLO9000: Better, faster, stronger,

    J. Redmon and A. Farhadi, “YOLO9000: Better, faster, stronger,” IEEE Conference on Computer Vision and Pattern Recognition , pp. 6517– 6525, 2016

    2016

  26. [34]

    YOLOv3: An incremental improvement,

    J. Redmon, “YOLOv3: An incremental improvement,” arXiv preprint arXiv:1804.02767, 2018

    2018 arXiv

  27. [35]

    YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

    C.-Y . Wang, A. Bochkovskiy, and H.-Y . M. Liao, “YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 7464–7475

    2023

  28. [36]

    YOLOX: Exceeding YOLO series in 2021,

    Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “YOLOX: Exceeding YOLO series in 2021,” ArXiv, vol. abs/2107.08430, 2021

    2021 arXiv

  29. [37]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in European conference on computer vision . Springer, 2020, pp. 213– 229

    2020

  30. [38]

    Lightweight thermal super-resolution and object detection for robust perception in adverse weather conditions,

    P. Shyam and H. Yoo, “Lightweight thermal super-resolution and object detection for robust perception in adverse weather conditions,” in IEEE/CVF Winter Conference on Applications of Computer Vision , 2024, pp. 7456–7467

    2024

  31. [39]

    Improved thermal infrared image super-resolution reconstruction method base on multimodal sensor fusion,

    Y . Jiang, Y . Liu, W. Zhan, and D. Zhu, “Improved thermal infrared image super-resolution reconstruction method base on multimodal sensor fusion,” Entropy, vol. 25, no. 6, 2023

    2023

  32. [40]

    Thermal object detection in difficult weather conditions using YOLO,

    M. Kri ˇsto, M. Ivasic-Kos, and M. Pobar, “Thermal object detection in difficult weather conditions using YOLO,” IEEE Access , vol. 8, pp. 125 459–125 476, 2020

    2020

  33. [41]

    Pheasant and C

    S. Pheasant and C. Haslegrave, Bodyspace: Anthropometry, Ergonomics and the Design of Work . CRC Press, 2018

    2018

  34. [42]

    Accessed: 2025-01-06

    Ultralytics repository. Accessed: 2025-01-06. [Online]. Available: https://github.com/ultralytics/ultralytics

    2025

  35. [43]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2018, pp. 7132–7141

    2018

  36. [44]

    Cbam: Convolutional block attention module,

    S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon, “Cbam: Convolutional block attention module,” in European Conference on Computer Vision , 2018, pp. 3–19

    2018

  37. [45]

    Spatial pyramid pooling in deep convolutional networks for visual recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Spatial pyramid pooling in deep convolutional networks for visual recognition,” IEEE Transactions on 12 Pattern Analysis and Machine Intelligence , vol. 37, no. 9, pp. 1904– 1916, 2015

    1904

  38. [46]

    CSPNet: A new backbone that can enhance learning capability of CNN,

    C.-Y . Wang, H.-Y . Mark Liao, Y .-H. Wu, P.-Y . Chen, J.-W. Hsieh, and I.- H. Yeh, “CSPNet: A new backbone that can enhance learning capability of CNN,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020, pp. 1571–1580

    2020

  39. [47]

    Survey on deep learning with class imbalance,

    J. Johnson and T. Khoshgoftaar, “Survey on deep learning with class imbalance,” Journal of Big Data , vol. 6, p. 27, 03 2019

    2019

  40. [48]

    Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection,

    X. Li, W. Wang, L. Wu, S. Chen, X. Hu, J. Li, J. Tang, and J. Yang, “Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection,” in Proceedings of the 34th Inter- national Conference on Neural Information Processing Systems , 2020

    2020

  41. [49]

    Enhancing geometric factors in model learning and inference for object detection and instance segmentation,

    Z. Zheng, P. Wang, D. Ren, W. Liu, R. Ye, Q. Hu, and W. Zuo, “Enhancing geometric factors in model learning and inference for object detection and instance segmentation,” IEEE Transactions on Cybernetics, vol. 52, no. 8, pp. 8574–8586, 2022

    2022

  42. [50]

    The regression analysis of binary sequences,

    D. R. Cox, “The regression analysis of binary sequences,” Journal of the royal statistical society series b-methodological , vol. 20, pp. 215–232, 1958

    1958

  43. [51]

    Layercam: Exploring hierarchical class activation maps for localization,

    P.-T. Jiang, C.-B. Zhang, Q. Hou, M.-M. Cheng, and Y . Wei, “Layercam: Exploring hierarchical class activation maps for localization,” IEEE Transactions on Image Processing , vol. 30, pp. 5875–5888, 2021. Xiao Ni received the B.Eng. degree from Tongji University and is current...

    2021

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.