arxiv: 2604.10419 · v1 · submitted 2026-04-12 · 💻 cs.ET

Recognition: unknown

Roadside LiDAR for Cooperative Safety Auditing at Urban Intersections: Toward Auditable V2X Infrastructure Intelligence

Bo Shang , Yiqiao Li

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:34 UTC · model grok-4.3

classification 💻 cs.ET

keywords roadside LiDARurban intersectionsnear-miss analysistrajectory constructionhuman-in-the-loop QAsafety auditingV2X cooperative perceptiontime-to-collision

0 comments

The pith

Roadside LiDAR with human quality assurance produces reliable near-miss evidence for auditing urban intersection safety.

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

This paper develops a framework that deploys roadside LiDAR at signalized urban intersections to build trajectories of vehicles and vulnerable road users, then applies iterative human review to clean the data before running near-miss calculations. The method matters because single-vehicle sensors suffer from occlusions and limited fields of view, leaving gaps in safety analysis that infrastructure sensing can fill. Using real data from a New York City intersection, the authors anchor their evaluation on a heavy-vehicle and bicycle encounter where direction-agnostic time-to-collision falls below one second while longitudinal measures stay above braking thresholds, exposing a lateral-intrusion conflict. Continuous-window testing across multiple rounds of quality assurance shows measurable drops in track fragmentation, spurious triggers, unstable geometry, and false cross-lane conflicts. The result supplies transportation agencies with a post-hoc auditing tool that generates defensible safety evidence without depending solely on crash records.

Core claim

The framework integrates LiDAR trajectory construction, iterative human-in-the-loop quality assurance, and interpretable near-miss analytics to yield auditable safety evidence at urban intersections. In the heavy-vehicle bicycle anchor case, direction-agnostic TTC drops below 1 s while longitudinal TTC remains above conservative thresholds, identifying lateral intrusion as the dominant conflict mechanism. Across broader continuous-window evaluation, multi-round QA analysis systematically reduces track fragmentation, spurious TTC triggers, unstable geometry, and cross-lane false conflicts, positioning roadside LiDAR as a practical post-hoc auditing mechanism for cooperative perception systems

What carries the argument

Iterative human-in-the-loop quality assurance applied to roadside LiDAR trajectories and direction-agnostic TTC metrics, which filters errors through repeated review rounds.

If this is right

Agencies can shift from crash-only safety reviews to proactive identification of high-risk interactions using infrastructure data.
Cooperative V2X systems gain a post-hoc validation layer that catches perception failures before they affect real-time decisions.
Lateral versus longitudinal conflict mechanisms become distinguishable, allowing targeted infrastructure changes such as adjusted signal timing or lane markings.
Statistical safety auditing scales to many intersections by processing continuous sensor streams rather than waiting for rare crash events.

Where Pith is reading between the lines

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

If the QA step can be partially automated while preserving audit logs, the framework could move from post-hoc to near-real-time use.
Pairing roadside LiDAR with additional sensors could address current limitations in weather or lighting that affect single-modality performance.
Applying the same continuous-window QA method to other intersection types would test whether the observed error reductions hold beyond the studied New York City site.

Load-bearing premise

The human reviewers and chosen TTC variants correctly identify all relevant safety errors and real collision risks without missing critical cases or introducing new distortions.

What would settle it

Independent comparison of the framework's flagged near-miss events against actual recorded crashes or data from alternative sensors at the same intersection showing no correlation or systematic mismatches.

Figures

Figures reproduced from arXiv: 2604.10419 by Bo Shang, Yiqiao Li.

**Figure 1.** Figure 1: Conceptual overview of the roadside LiDAR safety-auditing architecture. Raw point clouds are converted into detections, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Study intersection and roadside sensing setup used for [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 4.** Figure 4: Qualitative detection examples from the CenterPoint [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Spatiotemporal interaction window for the anchor heavy [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Anchor heavy vehicle–bicycle interaction. The x-axis [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

**Figure 7.** Figure 7: Post–8,000-frame vehicle–VRU near-miss overlay [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗

read the original abstract

Urban intersections expose the limitations of single-vehicle perception under occlusion and partial observability. In this study, we present an auditable roadside LiDAR framework for infrastructure-assisted safety analysis at a signalized urban intersection in New York City, developed and evaluated using real-world data. The proposed framework integrates trajectory construction, iterative human-in-the-loop quality assurance (QA), and interpretable near-miss analytics to produce defensible safety evidence from infrastructure sensing. Using a human-labeled heavy vehicle--bicycle interaction as an anchor case, we show that direction-agnostic time-to-collision (TTC) drops below 1s, while longitudinal TTC remains above conservative braking thresholds, revealing a lateral-intrusion-dominated conflict mechanism. Beyond individual cases, continuous-window evaluation and multi-round QA analysis demonstrate that the framework systematically reduces failure modes such as track fragmentation, spurious TTC triggers, unstable geometry, and cross-lane false conflicts. These results position roadside LiDAR as a practical post-hoc auditing mechanism for cooperative perception systems, with broader statistical validation discussed. This work provides a pathway toward scalable, data-driven safety auditing of urban intersections, enabling transportation agencies to identify and mitigate high-risk interactions beyond crash-based analyses.

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

This applies standard roadside LiDAR tracking and TTC metrics to a real NYC intersection with human QA for safety auditing, but stays descriptive without numbers or external checks.

read the letter

This paper takes roadside LiDAR data from a signalized intersection in New York and runs it through trajectory building plus iterative human review to flag near-miss events. The anchor example is a heavy vehicle and bicycle interaction where one TTC measure drops below one second while the longitudinal version stays above braking thresholds, which they interpret as a lateral conflict. They also describe how continuous-window checks and multiple QA rounds cut down on track breaks, false triggers, and cross-lane errors. That setup is a straightforward way to turn fixed-sensor output into something an agency could audit after the fact. The real-world data collection and focus on interpretable metrics are the parts that hold up. The limitation is that the improvements are shown only through description of one case and the QA process itself. No pre/post error counts, inter-reviewer agreement figures, or matches to actual crash logs or co-located radar appear, so it is hard to judge how much the human loop actually fixes versus how much it might introduce its own bias. The TTC thresholds are presented as conservative but without independent confirmation against outcomes. This is the sort of applied demonstration that could interest transportation planners or groups working on infrastructure perception. It does not introduce new algorithms or derivations, so its value is in showing the auditing workflow on live data. I would send it for peer review because the deployment context is concrete and the questions it raises about validation are answerable with more data; the authors should expect referees to ask for quantitative error analysis and ground-truth comparisons.

Referee Report

2 major / 1 minor

Summary. The paper presents a roadside LiDAR framework for post-hoc safety auditing at urban intersections, combining trajectory construction from infrastructure sensors, iterative human-in-the-loop quality assurance (QA), and interpretable near-miss metrics such as direction-agnostic and longitudinal time-to-collision (TTC). Using real-world data from a signalized New York City intersection, it analyzes an anchor case of heavy vehicle–bicycle interaction showing direction-agnostic TTC dropping below 1 s while longitudinal TTC stays above braking thresholds, and claims that continuous-window evaluation plus multi-round QA systematically reduces failure modes including track fragmentation, spurious TTC triggers, unstable geometry, and cross-lane false conflicts, positioning the approach as auditable infrastructure intelligence for V2X systems.

Significance. If the central claims hold after quantitative strengthening, the work offers a practical pathway for transportation agencies to generate defensible near-miss safety evidence from roadside LiDAR, extending beyond crash-based statistics to proactive auditing of cooperative perception systems. The emphasis on human-in-the-loop QA for auditability and the use of real-world multi-sensor data at an operational intersection are clear strengths that align with emerging V2X infrastructure needs.

major comments (2)

Abstract: The claim that 'continuous-window evaluation and multi-round QA analysis demonstrate that the framework systematically reduces failure modes such as track fragmentation, spurious TTC triggers, unstable geometry, and cross-lane false conflicts' is load-bearing for the central contribution yet rests solely on descriptive demonstration from one anchor case and human review, with no reported pre/post-QA quantitative error rates, inter-annotator agreement statistics, baseline comparisons, or statistical tests.
Abstract (anchor-case analysis): The interpretation that direction-agnostic TTC dropping below 1 s while longitudinal TTC remains above thresholds reveals a 'lateral-intrusion-dominated conflict mechanism' lacks independent validation against actual crash records, co-located camera/radar ground truth, or alternative TTC formulations, leaving open whether the chosen thresholds correlate with real collision risk.

minor comments (1)

Abstract: The phrase 'broader statistical validation discussed' is referenced but not elaborated with specifics on methods, datasets, or sections where such validation appears, which would aid readers in assessing the scope of the current results.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We are grateful for the referee's insightful comments, which help strengthen the presentation of our work on auditable roadside LiDAR for safety auditing. Below we respond to each major comment, indicating the revisions made to the manuscript.

read point-by-point responses

Referee: Abstract: The claim that 'continuous-window evaluation and multi-round QA analysis demonstrate that the framework systematically reduces failure modes such as track fragmentation, spurious TTC triggers, unstable geometry, and cross-lane false conflicts' is load-bearing for the central contribution yet rests solely on descriptive demonstration from one anchor case and human review, with no reported pre/post-QA quantitative error rates, inter-annotator agreement statistics, baseline comparisons, or statistical tests.

Authors: We acknowledge that the evidence for systematic reduction is primarily descriptive based on the anchor case and the QA process detailed in the methods. To address this, we have revised the abstract to use more precise language, stating that the framework 'illustrates reductions in failure modes through continuous-window evaluation and multi-round QA' in the context of the analyzed case. Additionally, we have added quantitative details in a new subsection, including counts of corrected tracks per QA round and observed decreases in spurious triggers, along with notes on the human review process. While a full statistical analysis across a larger corpus would require expanded data collection, these additions provide more concrete support for the claim within the scope of the current study. revision: partial
Referee: Abstract (anchor-case analysis): The interpretation that direction-agnostic TTC dropping below 1 s while longitudinal TTC remains above thresholds reveals a 'lateral-intrusion-dominated conflict mechanism' lacks independent validation against actual crash records, co-located camera/radar ground truth, or alternative TTC formulations, leaving open whether the chosen thresholds correlate with real collision risk.

Authors: The anchor case serves to demonstrate the framework's capability to generate interpretable near-miss evidence that can highlight conflict types like lateral intrusions, which may be missed by longitudinal-only metrics. The thresholds are selected based on established safety literature for conservative braking distances. We agree that direct validation with crash data would be ideal but is not available for this dataset, as crash records for the specific time period and location were not accessible. In the revised manuscript, we have added a limitations paragraph discussing this and included a brief sensitivity analysis comparing the direction-agnostic TTC with standard formulations to show robustness. This positions the work as a proof-of-concept for auditable auditing rather than a validated risk predictor. revision: yes

standing simulated objections not resolved

Direct validation of the near-miss metrics against actual crash records or co-located ground truth data.

Circularity Check

0 steps flagged

No circularity: empirical framework relies on direct data processing and human review

full rationale

The paper describes an infrastructure LiDAR pipeline for trajectory extraction, human-in-the-loop QA, and TTC-based near-miss analysis on real-world urban intersection data. No equations, fitted parameters, model predictions, or self-referential definitions appear in the provided text; the central claims rest on observed reductions in failure modes after multi-round QA rather than any derivation that reduces to its own inputs by construction. Self-citations, if present, are not load-bearing for the reported results, which are grounded in primary sensor data and annotator review rather than prior author theorems or ansatzes.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are required; the framework relies on standard trajectory construction and TTC definitions already present in traffic-safety literature.

pith-pipeline@v0.9.0 · 5512 in / 1147 out tokens · 29948 ms · 2026-05-10T16:34:51.399643+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 1 canonical work pages

[1]

Simple online and realtime tracking

Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In2016 IEEE International Conference on Image Processing (ICIP), pages 3464–3468, 2016

2016
[2]

Gettman and L

D. Gettman and L. Head. Surrogate safety mea- sures from traffic simulation models. Technical Report FHW A-RD-03-050, Federal Highway Ad- ministration (FHW A), 2003. HTML landing page: https://www.fhwa.dot.gov/publications/research/safety/03050/07.cfm

2003
[3]

Huang et al

T. Huang et al. Vehicle-to-everything cooperative perception for autonomous driving: A survey. arXiv preprint, 2023

2023
[4]

The swedish traffic conflicts technique

Christer Hyd ´en. The swedish traffic conflicts technique. Technical report / calibration study, 1987

1987
[5]

Phd dissertation, Link¨oping University, Link¨oping, Sweden, 2005

Jonas Jansson.Collision Avoidance Theory: with Applica- tion to Automotive Collision Mitigation. Phd dissertation, Link¨oping University, Link¨oping, Sweden, 2005

2005
[6]

Ji et al

Y . Ji et al. Toward autonomous vehicles: A survey on coop- erative vehicle-infrastructure system.iScience, 2024

2024
[7]

Johnsson et al

C. Johnsson et al. A relative approach to the validation of surrogate measures of safety.Accident Analysis & Preven- tion, 2021

2021
[8]

Lin et al

C. Lin et al. Vehicle detection and tracking with roadside lidar using low-latency edge computing.Sensors, 2023

2023
[9]

Mo et al

Y . Mo et al. Enhanced perception for autonomous vehicles at intersections using roadside sensing and v2x communica- tion.Sensors, 2024

2024
[10]

Urbaning-v2x: A large-scale multi-vehicle, multi- infrastructure dataset across multiple intersections for co- operative perception

Karthikeyan Chandra Sekaran, Markus Geisler, Do- minik R ¨oßle, Adithya Mohan, Daniel Cremers, Wolfgang Utschick, Michael Botsch, Werner Huber, and Torsten Sch¨on. Urbaning-v2x: A large-scale multi-vehicle, multi- infrastructure dataset across multiple intersections for co- operative perception. arXiv preprint, 2025. Code/dataset: https://github.com/thi-a...

2025
[11]

Ward, Gabriel Agamennoni, Stewart Worrall, Asher Bender, and Eduardo Nebot

James R. Ward, Gabriel Agamennoni, Stewart Worrall, Asher Bender, and Eduardo Nebot. Extending time to col- lision for probabilistic reasoning in general traffic scenar- ios.Transportation Research Part C: Emerging Technolo- gies, 51:66–82, 2015

2015
[12]

Yazgan et al

M. Yazgan et al. Collaborative perception datasets in au- tonomous driving. arXiv preprint, 2024

2024
[13]

Center- based 3d object detection and tracking

Tianwei Yin, Xingyi Zhou, and Philipp Kr ¨ahenb¨uhl. Center- based 3d object detection and tracking. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11784–11793, 2021

2021
[14]

Dair-v2x: A large-scale dataset for vehicle- infrastructure cooperative 3d object detection

Haibao Yu et al. Dair-v2x: A large-scale dataset for vehicle- infrastructure cooperative 3d object detection. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

2022
[15]

V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and fore- casting

Haibao Yu et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and fore- casting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

2023
[16]

Road- side lidar-based scene understanding toward intelligent traf- fic perception: A comprehensive review.ISPRS Journal of Photogrammetry and Remote Sensing, 2026

Jiaxing Zhang, Chengjun Ge, Wen Xiao, Miao Tang, Jon Mills, Benjamin Coifman, and Nengcheng Chen. Road- side lidar-based scene understanding toward intelligent traf- fic perception: A comprehensive review.ISPRS Journal of Photogrammetry and Remote Sensing, 2026

2026
[17]

Tumtraf v2x cooperative perception dataset

Walter Zimmer et al. Tumtraf v2x cooperative perception dataset. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. Project page: https://tum-traffic-dataset.github.io/tumtraf- v2x/ ; arXiv: https://arxiv.org/abs/2403.01316

work page arXiv 2024