arxiv: 2604.25887 · v1 · submitted 2026-04-28 · 💻 cs.CV · cs.AI· cs.RO· cs.SY· eess.SY

Recognition: unknown

No Pedestrian Left Behind: Real-Time Detection and Tracking of Vulnerable Road Users for Adaptive Traffic Signal Control

Anas Gamal Aly , Hala ElAarag

Authors on Pith no claims yet

Pith reviewed 2026-05-07 16:37 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.ROcs.SYeess.SY

keywords vulnerable road usersadaptive traffic signalsobject detectionpedestrian trackingYOLOv12Monte Carlo simulationtraffic signal controlVRU safety

0 comments

The pith

Real-time detection and tracking extends traffic signals to prevent vulnerable road users from being stranded mid-crossing.

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

Fixed pedestrian signals use preset timings that can leave elderly, disabled, or distracted people stranded when the light changes before they finish crossing. The paper introduces NPLB, a system that detects and tracks these vulnerable road users in real time and automatically extends the signal phase if the remaining time falls below a safety threshold. It combines a fine-tuned YOLOv12 detector with ByteTrack tracking and a simple adaptive controller. In 10,000 Monte Carlo simulations, the approach reduces stranding rates from 9.10 percent to 2.60 percent, a 71.4 percent safety gain, while triggering extensions in only 12.1 percent of cycles. The work shows that modest extensions to existing vision models can address a practical safety gap without major changes to traffic flow.

Core claim

NPLB monitors vulnerable road users in crosswalks with fine-tuned YOLOv12 detection and ByteTrack multi-object tracking, then applies an adaptive controller that extends the pedestrian phase when remaining time drops below a critical threshold. Across 10,000 Monte Carlo simulations, this produces a 71.4 percent improvement in VRU safety by lowering stranding rates from 9.10 percent to 2.60 percent, while requiring signal extensions in only 12.1 percent of crossing cycles.

What carries the argument

The adaptive controller that tracks remaining crossing time for detected vulnerable road users and triggers signal extensions when time falls below a safety threshold.

If this is right

Stranding rates for vulnerable road users fall from 9.10% to 2.60% under the simulated conditions.
Signal extensions occur in only 12.1% of crossing cycles, limiting disruption to vehicle flow.
YOLOv12 reaches 0.756 mAP@0.5 on the BGVP dataset, outperforming the other four tested detectors.
The full pipeline runs on standard computer vision components without requiring new hardware beyond cameras.

Where Pith is reading between the lines

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

Cities could retrofit existing signal controllers with the detection module to improve safety at high-VRU intersections.
The low extension rate suggests the method could be combined with vehicle-priority logic without creating major delays.
Performance gaps in adverse weather would need direct measurement before widespread rollout.

Load-bearing premise

The object detection and tracking models maintain reliable performance in real-world conditions with varying lighting, weather, occlusions, and camera angles, and the Monte Carlo parameters accurately reflect actual pedestrian behavior and traffic dynamics.

What would settle it

A controlled field deployment at a real intersection that measures the actual percentage of vulnerable road users stranded before versus after NPLB installation.

Figures

Figures reproduced from arXiv: 2604.25887 by Anas Gamal Aly, Hala ElAarag.

**Figure 1.** Figure 1: NPLB Architecture The NPLB system addresses a critical gap in pedestrian safety by dynamically adapting traffic signal timing based on real-time detection of vulnerable road users. The system architecture, illustrated in view at source ↗

**Figure 2.** Figure 2: Heatmap of Parameter Sweep Showing Stranding view at source ↗

**Figure 3.** Figure 3: Comparison of Stranding Rates Between Fixed view at source ↗

**Figure 4.** Figure 4: Frequency Distribution of Signal Extensions Trig view at source ↗

**Figure 5.** Figure 5: Box Plot Comparing Pedestrian Signal Durations view at source ↗

**Figure 6.** Figure 6: Training and Validation Metrics for YOLOv5 Across 500 Epochs, Showing Convergence of Box Loss, Objectness Loss, view at source ↗

**Figure 7.** Figure 7: Training and Validation Metrics for YOLOv11 Across 500 Epochs, Showing Convergence of Box Loss, Objectness Loss, view at source ↗

**Figure 8.** Figure 8: Training and Validation Metrics for YOLOv12 Across 500 Epochs, Showing Convergence of Box Loss, Objectness Loss, view at source ↗

read the original abstract

Current pedestrian crossing signals operate on fixed timing without adjustment to pedestrian behavior, which can leave vulnerable road users (VRUs) such as the elderly, disabled, or distracted pedestrians stranded when the light changes. We introduce No Pedestrian Left Behind (NPLB), a real-time adaptive traffic signal system that monitors VRUs in crosswalks and automatically extends signal timing when needed. We evaluated five state-of-the-art object detection models on the BGVP dataset, with YOLOv12 achieving the highest mean Average Precision at 50% (mAP@0.5) of 0.756. NPLB integrates our fine-tuned YOLOv12 with ByteTrack multi-object tracking and an adaptive controller that extends pedestrian phases when remaining time falls below a critical threshold. Through 10,000 Monte Carlo simulations, we demonstrate that NPLB improves VRU safety by 71.4%, reducing stranding rates from 9.10% to 2.60%, while requiring signal extensions in only 12.1% of crossing cycles.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper integrates standard YOLOv12 detection and ByteTrack tracking into a threshold-based adaptive signal controller and reports large safety gains from 10,000 Monte Carlo runs, but those gains rest on unvalidated simulation parameters.

read the letter

The main thing to know is that this work takes off-the-shelf models, fine-tunes YOLOv12 on the BGVP dataset to reach mAP@0.5 of 0.756, wires it to ByteTrack, and adds a simple rule that extends the pedestrian phase when remaining time drops below a threshold. The headline result is a simulated drop in stranding from 9.1% to 2.6% with extensions needed in only 12% of cycles. That pipeline is new in its exact combination and dataset choice, even if the pieces themselves are not novel. The authors also ran a clean comparison of five detectors and kept the controller lightweight, which is useful engineering detail. The low extension rate is a practical plus that shows the system would not constantly disrupt traffic flow. The soft spots sit in the evaluation. The Monte Carlo setup uses distributions for crossing speed, arrival, and distraction that are never tied to field measurements or existing traffic studies, so it is unclear whether the baseline 9.1% stranding rate matches real intersections. Detection and tracking errors are also not injected into the loop, even though mAP of 0.756 implies misses and false positives that would affect controller decisions. Without those checks or any real-world video or deployment data, the 71% safety figure stays hypothetical. This paper is for people working on CV applications in transportation or traffic-signal engineers who want a concrete starting point for adaptive VRU handling. It is not a methods paper but a system demonstration. I would send it to peer review; the referees can push for sensitivity analysis on the simulation parameters and some form of error modeling, which would make the claims more credible.

Referee Report

3 major / 2 minor

Summary. The paper proposes No Pedestrian Left Behind (NPLB), a real-time adaptive traffic signal control system that uses object detection and multi-object tracking to monitor vulnerable road users (VRUs) in crosswalks and extend pedestrian phase timings when remaining time falls below a threshold. Five state-of-the-art detectors are evaluated on the BGVP dataset, with fine-tuned YOLOv12 achieving the highest mAP@0.5 of 0.756; this is integrated with ByteTrack. The central empirical claim is that 10,000 Monte Carlo simulations of the adaptive controller reduce VRU stranding rates from 9.10% to 2.60% (71.4% improvement) while requiring signal extensions in only 12.1% of cycles.

Significance. If the reported safety gains hold under realistic conditions, the work could contribute to practical adaptive signal systems that improve VRU safety with limited disruption to vehicular flow. The combination of recent detection/tracking models with a simple threshold-based controller is a straightforward application of computer vision to transportation. The low reported extension frequency is a strength for traffic efficiency. However, the moderate detection accuracy and exclusive reliance on unvalidated simulations limit the immediate significance and generalizability of the 71.4% figure.

major comments (3)

[Evaluation / Monte Carlo simulations] The Monte Carlo simulation parameters (VRU crossing speeds, arrival processes, distraction durations, and remaining-time thresholds) are not stated to have been derived from or validated against real intersection field data or observational studies. Consequently the baseline stranding rate of 9.10% and the 71.4% reduction cannot be assessed for realism.
[System description and evaluation] Detection and tracking errors are not injected into the controller loop despite the reported YOLOv12 mAP@0.5 of 0.756 (which implies non-negligible miss and false-positive rates). The simulation therefore assumes perfect VRU localization, which likely overstates the safety benefit.
[Abstract and results] No real-world deployment, on-site testing, or sensitivity analysis on key parameters is presented. All safety claims rest solely on the 10,000 simulation runs without error propagation or comparison to observed stranding rates at actual sites.

minor comments (2)

The BGVP dataset size, diversity, camera angles, and annotation protocol are not described, making it difficult to interpret the mAP@0.5 result.
Clarify whether the critical remaining-time threshold is a fixed hyper-parameter or learned/adapted per intersection.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their constructive comments on our manuscript. We have carefully reviewed each major concern and provide point-by-point responses below. Where revisions are feasible, we have incorporated or will incorporate changes to strengthen the paper while maintaining the integrity of our simulation-based evaluation.

read point-by-point responses

Referee: [Evaluation / Monte Carlo simulations] The Monte Carlo simulation parameters (VRU crossing speeds, arrival processes, distraction durations, and remaining-time thresholds) are not stated to have been derived from or validated against real intersection field data or observational studies. Consequently the baseline stranding rate of 9.10% and the 71.4% reduction cannot be assessed for realism.

Authors: We appreciate this observation. The simulation parameters were drawn from standard values in the transportation literature (e.g., pedestrian walking speeds of 1.0–1.4 m/s for vulnerable users and typical distraction durations from observational studies), but explicit sourcing and justification were not provided in the original text. In the revised manuscript we will add a dedicated subsection citing the literature sources for each parameter and include a sensitivity analysis that varies crossing speeds, arrival rates, and thresholds to demonstrate the robustness of the reported 71.4% reduction. revision: yes
Referee: [System description and evaluation] Detection and tracking errors are not injected into the controller loop despite the reported YOLOv12 mAP@0.5 of 0.756 (which implies non-negligible miss and false-positive rates). The simulation therefore assumes perfect VRU localization, which likely overstates the safety benefit.

Authors: This is a fair critique. The current simulations use idealized VRU positions to isolate the adaptive controller's contribution. We agree that this assumption may overstate benefits given the detector performance. We will revise the evaluation to include an error-injection experiment that randomly suppresses detections and introduces false positives at rates consistent with the precision/recall observed on the BGVP dataset, thereby providing a more conservative estimate of safety gains under realistic detection conditions. revision: yes
Referee: [Abstract and results] No real-world deployment, on-site testing, or sensitivity analysis on key parameters is presented. All safety claims rest solely on the 10,000 simulation runs without error propagation or comparison to observed stranding rates at actual sites.

Authors: We acknowledge that the work is entirely simulation-based and that real-world deployment or on-site testing was not performed. This is an inherent scope limitation of the current study. In the revision we will add sensitivity analysis across key parameters and a basic error-propagation discussion. However, direct comparison to observed stranding rates at instrumented intersections and full on-site validation require field data collection and regulatory approvals that are outside the present scope; we have expanded the limitations and future-work sections to address this explicitly. revision: partial

standing simulated objections not resolved

Real-world deployment, on-site testing, and direct comparison against observed stranding rates at actual intersections, as these data were not collected in the current simulation-focused study.

Circularity Check

0 steps flagged

Monte Carlo simulation outputs are independent of detection model fitting

full rationale

The paper's central result (71.4% safety improvement) is produced by running 10,000 independent Monte Carlo trials of the adaptive controller. Detection performance is measured on the external BGVP dataset and yields a fixed mAP value that is then used as input to the simulator; the simulator parameters for crossing speeds, arrival processes, and thresholds are stated separately and not fitted to the same data that produces the headline percentages. No equation equates the reported stranding rates back to the YOLOv12 weights or to any self-citation. The derivation chain therefore terminates in external simulation runs rather than reducing to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central performance claims rest on unstated simulation assumptions about pedestrian speeds, arrival distributions, and detection reliability under real conditions; no free parameters or invented entities are explicitly introduced in the abstract.

pith-pipeline@v0.9.0 · 5503 in / 1184 out tokens · 41216 ms · 2026-05-07T16:37:40.880887+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

33 extracted references · 15 canonical work pages · 1 internal anchor

[1]

Ahmed S Abdelrahman, Mohamed Abdel-Aty, and Dongdong Wang. 2025. Video- To-Text Pedestrian Monitoring (VTPM): Leveraging Large Language Models for Privacy-Preserve Pedestrian Activity Monitoring at Intersections. InProceedings of the Winter Conference on Applications of Computer Vision. 366–375

2025
[2]

Ahmed S Abdelrahman, Zubayer Islam, and Mohamed Abdel-Aty. 2025. VRU- CrossSafe for Crossing Intention Prediction of Vulnerable Road Users for Im- proving Safe Crossing at Intersections.npj Sustainable Mobility and Transport2, 1 (2025), 20

2025
[3]

Md Jamil Ahsan, Mohamed Abdel-Aty, and Ahmed S Abdelrahman. 2025. Evalu- ating the Safety Impact of Mid-Block Pedestrian Signals (MPS).Accident Analysis & Prevention210 (2025), 107847

2025
[4]

Gorkem Akyol, Ismet Goksad Erdagi, Mehmet Ali Silgu, and Hilmi Berk Celikoglu
[5]

https://doi

Adaptive Signal Control to Enhance Effective Green Times for Pedestrians: A Case Study.Transportation Research Procedia47 (2020), 704–711. https://doi. org/10.1016/j.trpro.2020.03.150 22nd EURO Working Group on Transportation Meeting, EWGT 2019, 18th – 20th September 2019, Barcelona, Spain

work page doi:10.1016/j.trpro.2020.03.150 2020
[6]

Raul Avelar, Eun Sug Park, Karen Dixon, Xiao Li, Minh Li, and Bahar Dadashova
[7]

https://www.fhwa.dot.gov/publications/research/safety/20072/20072

Developing Crash Modification Factors for Adaptive Signal Control Tech- nologies. https://www.fhwa.dot.gov/publications/research/safety/20072/20072. pdf
[8]

Alben Rome Bagabaldo and Jürgen Hackl. 2025. Improving Pedestrian Safety at Intersections Using Probabilistic Models and Monte Carlo Simulations.arXiv preprint arXiv:2503.07805(2025)

work page arXiv 2025
[9]

Manuel Burgos-Payán, Francisco-Javier Correa-Moreno, and Jesús-Manuel Riquelme-Santos. 2012. Improving the Energy Efficiency of Street Lighting: A Case in the South of Spain. In2012 9th International Conference on the European Energy Market. IEEE, 1–8

2012
[10]

Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. 2023. Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR). 9686–9696. https://openaccess.thecvf.com/content/CVPR2023/papers/Cao_Observation- Centric_SORT_Rethinking...

2023
[11]

Colorado Department of Transportation. 2024. CDOT Pedestrian Signal Timing Guidelines. https://www.codot.gov/safety/traffic-safety/assets/documents

2024
[12]

Piotr Dollar, Christian Wojek, Bernt Schiele, and Pietro Perona. 2009. Pedestrian Detection: A Benchmark. In2009 IEEE Conference on Computer Vision and Pattern Recognition. 304–311. https://doi.org/10.1109/CVPR.2009.5206631

work page doi:10.1109/cvpr.2009.5206631 2009
[13]

Divya Gangwani and Pranav Gangwani. 2021. Applications of Machine Learn- ing and Artificial Intelligence in Intelligent Transportation System: A Review. Applications of Artificial Intelligence and Machine Learning: Select Proceedings of ICAAAIML 2020(2021), 203–216

2021
[14]

Zhi Huang, Xiangyi Liu, Xiaolin Song, and Yin He. 2017. Probabilistic Risk Assessment for Pedestrian-Vehicle Collision Considering Uncertainties of Pedes- trian Mobility.Traffic Injury Prevention18, 6 (2017), 650–656. https://doi.org/10. 1080/15389588.2017.1283026 arXiv:https://doi.org/10.1080/15389588.2017.1283026 PMID: 28112561

work page doi:10.1080/15389588.2017.1283026 2017
[15]

Drame, Mamadou L

Abdoulaye Kama, Moussa Diallo, Mamadou S. Drame, Mamadou L. Ndiaye, Ababacar Ndiaye, and Pape A. Ndiaye. 2017. Monitoring the Performance of Solar Street Lights in Sahelian Environment: Case Study of Senegal. In2017 10th International Conference on Developments in eSystems Engineering (DeSE). 56–61. https://doi.org/10.1109/DeSE.2017.43

work page doi:10.1109/dese.2017.43 2017
[16]

Marta Kolasa. 2016. The Concept of Intelligent System for Street Lighting Control Using Artificial Neural Networks.Prz. Elektrotechniczny92, 7 (2016), 32–37

2016
[17]

Iuliia Kotseruba, Amir Rasouli, and John K. Tsotsos. 2016. Joint Attention in Autonomous Driving (JAAD) Dataset. https://arxiv.org/abs/1609.04741

work page arXiv 2016
[18]

Lawrence Zitnick, and Piotr Dollár

Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dollár
[19]

Microsoft COCO: Common Objects in Context

Microsoft COCO: Common Objects in Context. arXiv:1405.0312 [cs.CV] https://arxiv.org/abs/1405.0312

work page internal anchor Pith review arXiv
[20]

Prabu Mohandas, Jerline Sheebha Anni Dhanaraj, and Xiao-Zhi Gao. 2019. Arti- ficial Neural Network Based Smart and Energy Efficient Street Lighting System: A Case Study for Residential Area in Hosur.Sustainable Cities and Society48 (2019), 101499. https://doi.org/10.1016/j.scs.2019.101499

work page doi:10.1016/j.scs.2019.101499 2019
[21]

Moran and Debra F

Marcel E. Moran and Debra F. Laefer. 2024. Multiscale Analysis of Pedestrian Crossing Distance.Journal of the American Planning Asso- ciation0, 0 (2024), 1–15. https://doi.org/10.1080/01944363.2024.2394610 arXiv:https://doi.org/10.1080/01944363.2024.2394610

work page doi:10.1080/01944363.2024.2394610 2024
[22]

Deepti Muley, Wael Alhajyaseen, and Mohamed Kharbeche. 2017. An Overview of Pedestrian Signal Setting and Implementation in the State of Qatar.Procedia Computer Science109 (2017), 545–552. https://doi.org/10.1016/j.procs.2017.05.337 8th International Conference on Ambient Systems, Networks and Technologies, ANT-2017 and the 7th International Conference on...

work page doi:10.1016/j.procs.2017.05.337 2017
[23]

Amir Rasouli, Iuliia Kotseruba, and John K. Tsotsos. 2019. PIE: A Large-Scale Dataset and Models for Pedestrian Intention Estimation. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 6262–6271. https://openaccess.thecvf.com/content_ICCV_2019/papers/Rasouli_PIE_A_ Large-Scale_Dataset_and_Models_for_Pedestrian_Intention_Esti...

2019
[24]

Tobias Schlosser, Michael Friedrich, Trixy Meyer, and Danny Kowerko. 2024. A Consolidated Overview of Evaluation and Performance Metrics for Machine Learning and Computer Vision.Tobias Schlosser, Michael Friedrich, Trixy Meyer, and Danny Kowerko–Junior Professorship of Media Computing, Chemnitz Univer- sity of Technology9107 (2024)

2024
[25]

Devansh Sharma, Tihitina Hade, and Qing Tian. 2022. Comparison of Deep Object Detectors on a New Vulnerable Pedestrian Dataset.arXiv preprint arXiv:2212.06218(2022)

work page arXiv 2022
[26]

Asis Kumar Tripathy, Alekha Kumar Mishra, and Tapan Kumar Das. 2017. Smart Lighting: Intelligent and Weather Adaptive Lighting in Street Lights Using IoT. In2017 International Conference on Intelligent Computing, Instrumentation and Control Technologies (ICICICT). 1236–1239. https://doi.org/10.1109/ICICICT1. 2017.8342746

work page doi:10.1109/icicict1 2017
[27]

US Census Bureau. 2025. Older Adults Outnumber Children in 11 States and Nearly Half of U.S. Counties. https://www.census.gov/newsroom/press-releases/ 2025/older-adults-outnumber-children.html

2025
[28]

Department of Transportation

U.S. Department of Transportation. 2016. Lesson 8 - Federal Highway Adminis- tration University Course on Bicycle and Pedestrian Transportation, July 2006 - FHWA-HRT-05-100. https://www.fhwa.dot.gov/publications/research/safety/ pedbike/05085/pptchapt8.cfm

2016
[29]

Department of Transportation

U.S. Department of Transportation. 2024. Secretary Buttigieg Announces Proposed Rule to Ensure Passengers Who Use Wheelchairs Can Fly with Dignity. https://www.transportation.gov/briefing-room/secretary-buttigieg- announces-proposed-rule-ensure-passengers-who-use-wheelchairs-can

2024
[30]

Xinhai Wang and Wenhua Shao. 2025. Networked Sensor-Based Adaptive Traffic Signal Control for Dynamic Flow Optimization.Sensors25, 11 (2025), 3501. https://doi.org/10.3390/s25113501

work page doi:10.3390/s25113501 2025
[31]

Nicolai Wojke, Alex Bewley, and Dietrich Paulus. 2017. Simple Online and Real- time Tracking with a Deep Association Metric.arXiv preprint arXiv:1703.07402 (2017). https://arxiv.org/abs/1703.07402

work page arXiv 2017
[32]

World Health Organization. 2023. Pedestrian Safety: A Road Safety Manual for Decision-Makers and Practitioners, 2nd Edition. https://www.who.int/ publications/i/item/9789240072497. Accessed 2025-09-18

work page arXiv 2023
[33]

Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. 2022. ByteTrack: Multi-Object Tracking by Associating Every Detection Box. InEuropean Conference on Computer Vision. Springer, 1–21. Figure 6: Training and Validation Metrics for YOLOv5 Across 500 Epochs, Showing Convergence of Box Loss, Object...

2022