arxiv: 2605.00050 · v1 · submitted 2026-04-29 · 💻 cs.LG · cs.CV

Recognition: unknown

Learning physically grounded traffic accident reconstruction from public accident reports

Yanchen Guan , Haicheng Liao , Chengyue Wang , Zhenning Li

Authors on Pith no claims yet

Pith reviewed 2026-05-09 19:55 UTC · model grok-4.3

classification 💻 cs.LG cs.CV

keywords traffic accident reconstructionmultimodal learningpublic reportsphysical groundingroad topologycollision dynamicspre-impact motionreconstruction framework

0 comments

The pith

Public accident reports contain enough detail for learning models to reconstruct traffic crashes with physical accuracy.

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

The paper shows that routine textual accident reports can be turned into quantitative, physically grounded reconstructions of crashes by mapping their language onto road layouts, vehicle properties, and collision sequences. This matters because expert scene measurements and detailed investigations are too scarce and expensive to scale, leaving most real-world incidents unusable for safety analysis or training. The authors build a dataset of thousands of cases and a multimodal framework that first aligns report semantics with geometry, then generates consistent pre-impact paths, and finally refines interactions through local spatial and timing constraints. When tested, the resulting reconstructions match observed outcomes more closely than standard baselines across accuracy and consistency metrics.

Core claim

Accident reconstruction is posed as a parameterized multimodal learning task that takes public reports together with available scene measurements and produces outputs grounded in road topology, participant attributes, lane-consistent pre-impact motion, and collision dynamics. The framework achieves this grounding through semantic alignment, motion reconstruction, and localized geometric reasoning with temporal allocation. On a curated collection of more than six thousand real-world cases the method records stronger overall fidelity than representative baselines, with measurable gains in accident-point accuracy and collision consistency.

What carries the argument

A multimodal reconstruction framework that grounds textual report semantics to road topology and participant attributes, then reconstructs lane-consistent pre-impact motion and refines collision interactions via localized geometric reasoning and temporal allocation.

If this is right

Reconstruction fidelity improves on accident-point accuracy and collision consistency relative to prior methods.
Routine public reports become usable inputs for large-scale traffic safety studies.
Reconstructed scenarios supply data for simulation environments and autonomous-vehicle testing.

Where Pith is reading between the lines

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

Thousands of such automated reconstructions could surface recurring accident patterns that guide road-design changes.
The same grounding pipeline might be applied to other textual descriptions that need physical consistency, such as incident logs in robotics or insurance claims.
Integration with existing simulators could generate rare-event training data at low additional cost.

Load-bearing premise

Textual reports contain enough semantic detail to be reliably mapped onto physical road topology, participant attributes, and collision dynamics without extra expert measurements.

What would settle it

Compare model-generated trajectories and impact points against a held-out set of accidents that include independent laser-scanned scene data or expert-verified measurements; consistent large discrepancies in positions, speeds, or timings would disprove the claim.

Figures

Figures reproduced from arXiv: 2605.00050 by Chengyue Wang, Haicheng Liao, Yanchen Guan, Zhenning Li.

**Figure 1.** Figure 1: The input representations and corresponding input examples derived from the original accident report. In the same traffic accident report, scene factors are global attributes, while vehicle factors are individual attributes of the accident participants. of accident reports, in which global scene descriptions and vehicle-specific records are provided separately but jointly constrain accident evolution. The … view at source ↗

**Figure 2.** Figure 2: The preprocessing of velocity and trajectory data. Sparse trajectories and EDR records in the original accident reports are transformed into temporally dense supervision through alignment, fitting, and reconstruction, providing structured training targets for accident reconstruction. fitted keypoints, and trajectory evidence across accident cases under a unified spatial convention. To support scene-grounde… view at source ↗

**Figure 3.** Figure 3: Overall framework of our physically grounded accident reconstruction model. The overall architecture of the proposed model follows an encoder–decoder paradigm. The input scene is first encoded into a latent representation space. Based on this representation, a graph structure is constructed to model interactions among multiple traffic participants. The decoder then generates candidate trajectories from the… view at source ↗

**Figure 4.** Figure 4: Controllable visual reconstruction of traffic accident scenes. The framework first uses traffic simulation to preserve the physical consistency of scene geometry, vehicle motion and participant interactions, and then applies AI-based video generation to translate simulator outputs into realistic accident visualizations with matched environmental conditions and post-collision appearance. 3.9. Controllable v… view at source ↗

**Figure 5.** Figure 5: Controllable visualization results of a reconstructed traffic accident case, including multi-view pre-impact replay, postimpact scene extension, and rendering under different environmental conditions [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: Data distribution of key factors in the CISS-REC dataset. Yanchen Guan et al.: Preprint submitted to Elsevier Page 12 of 22 [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗

**Figure 7.** Figure 7: Metric landscape of accident reconstruction performance across baselines. Higher scores represent better performance across trajectory, collision, behavior, and geometry metrics [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗

**Figure 8.** Figure 8: Visualization results of representative accident cases. The dark solid lines denote the inferred trajectories generated by the proposed model, while the light dashed lines indicate the ground-truth trajectories recorded during the actual traffic accidents. trajectory reconstruction. To better simulate the real-world scenario of non-deep accident investigation, we injected random Gaussian noise with a maxim… view at source ↗

**Figure 9.** Figure 9: Visualization of an external traffic accident report. Some of the input information is inferred from the original accident report using Qwen3. To avoid privacy breaches, we have blurred the specific coordinates. architectures. All models use the same input, training process, and supervision. It is worth noting that PC-Crash and the Momentum-Energy, as classical baselines specifically designed for collisio… view at source ↗

**Figure 10.** Figure 10: Sensitivity of reconstruction metrics to random report-entry missingness. model to learn the temporal organization of trajectory evolution, including speed distribution, motion rhythm, and the timing structure leading to collision events. After removing this module, the inferred velocity error increases substantially, while trajectory-related errors slightly decrease. This observation further confirms t… view at source ↗

read the original abstract

Traffic accidents are routinely documented in textual reports, yet physically grounded accident reconstruction remains difficult because detailed scene measurements and expert reconstructions are scarce, costly and hard to scale. Here we formulate accident reconstruction from publicly accessible reports and scene measurements as a parameterized multimodal learning problem. We construct CISS-REC, a dataset of 6,217 real-world accident cases curated from the NHTSA Crash Investigation Sampling System, and develop a reconstruction framework that grounds report semantics to road topology and participant attributes, reconstructs lane consistent pre-impact motion, and refines collision relevant interactions through localized geometric reasoning and temporal allocation. Our method outperforms representative baselines on CISS-REC, achieving the strongest overall reconstruction fidelity, including improved accident point accuracy and collision consistency. These results show that public accident reports can serve as scalable computational substrates for quantitatively verifiable accident reconstruction, with potential value for traffic safety analysis, simulation and autonomous driving research.

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 builds a new dataset from NHTSA reports and a multimodal pipeline for text-to-physics accident reconstruction, but the evidence for reliable grounding from text alone is still thin.

read the letter

The main thing here is that they've created CISS-REC with 6,217 real cases pulled from public NHTSA reports and paired it with a framework that grounds report text to road topology and attributes, then reconstructs lane-consistent pre-impact motion and refines collisions via localized geometric reasoning. That specific combination for scaling reconstructions from text is new in the cited literature and directly targets the data scarcity problem in traffic safety work. They report stronger fidelity and collision consistency than baselines on their data, which is a practical result worth noting. The curation step itself is useful because it turns abundant textual records into something that can be tested against physical quantities. The soft spots are the missing specifics. The abstract and high-level description give no architecture, loss functions, metrics, or data split details, so it's hard to tell how much the gains come from the report semantics versus the added geometric steps. The risk that ambiguities like missing speeds or angles get resolved by dataset priors rather than the text is real, and the circularity concern holds if verification isn't cleanly separated from the fitted model. The stress-test note on whether reports contain enough quantitative detail is the central open question. This is for people doing large-scale safety analysis, simulation, or AV data augmentation who need more volume than expert reconstructions allow. A reader focused on multimodal grounding for physical tasks could extract ideas from the pipeline even before the numbers are fully vetted. I would send it for peer review so the methods and verification steps can be checked in detail.

Referee Report

3 major / 1 minor

Summary. The paper formulates traffic accident reconstruction from public textual reports as a parameterized multimodal learning problem. It introduces the CISS-REC dataset of 6,217 curated NHTSA cases and presents a framework that grounds report semantics to road topology and participant attributes, reconstructs lane-consistent pre-impact motion, and applies localized geometric reasoning plus temporal allocation for collision refinement. The method is reported to outperform representative baselines on CISS-REC in reconstruction fidelity, accident point accuracy, and collision consistency, supporting the claim that public reports can serve as scalable substrates for quantitatively verifiable accident reconstruction.

Significance. If the central claim holds with rigorous validation, the work would be significant for traffic safety analysis, simulation, and autonomous driving research by demonstrating a scalable, data-driven alternative to scarce expert reconstructions. The use of a large public dataset and emphasis on physical consistency could enable broader quantitative studies of accident dynamics that are currently limited by data availability.

major comments (3)

[Abstract] Abstract: The claim of outperformance on reconstruction fidelity and collision consistency is presented without any description of the model architecture, loss functions, evaluation metrics, data splits, or controls for selection bias in the 6,217 cases. This omission is load-bearing for assessing whether the results support the central claim of quantitatively verifiable reconstruction from textual reports.
[Abstract] Abstract and method description: The parameterized multimodal learning formulation risks circularity because performance metrics appear tied to quantities derived from the same fitted model on CISS-REC without explicit separation between training and independent physical verification steps. This directly affects the 'quantitatively verifiable' claim, as grounding and refinement may rely on learned priors rather than explicit report content.
[Abstract] Abstract: The framework is described as grounding report semantics to road topology, participant attributes, and collision dynamics, yet no indication is given of how common ambiguities (missing speeds, imprecise angles, omitted lane widths) are resolved without external measurements. This is central to the weakest assumption that textual reports contain sufficient quantitative detail for verifiable physical grounding.

minor comments (1)

[Abstract] Abstract: The phrase 'localized geometric reasoning and temporal allocation' is introduced without definition or reference to a specific section, reducing clarity of the method overview.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments on our work. We address each of the major comments in detail below, providing clarifications and indicating revisions to the manuscript where appropriate.

read point-by-point responses

Referee: [Abstract] Abstract: The claim of outperformance on reconstruction fidelity and collision consistency is presented without any description of the model architecture, loss functions, evaluation metrics, data splits, or controls for selection bias in the 6,217 cases. This omission is load-bearing for assessing whether the results support the central claim of quantitatively verifiable reconstruction from textual reports.

Authors: The abstract is intentionally brief to highlight the core contribution. Detailed descriptions of the model architecture (Section 3.1), loss functions (Section 3.4), evaluation metrics (Section 4.3), data splits (Section 4.1), and selection bias controls (Section 2.3) are provided in the main text. To improve accessibility, we will revise the abstract to include a short summary of the evaluation protocol and key metrics used. revision: yes
Referee: [Abstract] Abstract and method description: The parameterized multimodal learning formulation risks circularity because performance metrics appear tied to quantities derived from the same fitted model on CISS-REC without explicit separation between training and independent physical verification steps. This directly affects the 'quantitatively verifiable' claim, as grounding and refinement may rely on learned priors rather than explicit report content.

Authors: We believe there is no circularity in our approach. The training process uses textual reports and basic scene attributes to learn the grounding and motion reconstruction. The evaluation metrics, however, incorporate independent physical verification steps, including forward simulation of the reconstructed trajectories using Newtonian mechanics and comparison against available quantitative measurements in the CISS dataset that were not used in training. We will add explicit discussion of this separation in the revised manuscript to strengthen the 'quantitatively verifiable' claim. revision: partial
Referee: [Abstract] Abstract: The framework is described as grounding report semantics to road topology, participant attributes, and collision dynamics, yet no indication is given of how common ambiguities (missing speeds, imprecise angles, omitted lane widths) are resolved without external measurements. This is central to the weakest assumption that textual reports contain sufficient quantitative detail for verifiable physical grounding.

Authors: The resolution of ambiguities is addressed through the multimodal grounding module, which combines semantic parsing of the report with probabilistic inference over possible values constrained by road topology and participant attributes (detailed in Section 3.2). For instance, missing speeds are inferred from contextual descriptions (e.g., 'high speed') mapped to distributions, then refined by collision dynamics. We will include a new paragraph in the method section with concrete examples of ambiguity handling to clarify this process. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper constructs CISS-REC from NHTSA real-world cases and frames reconstruction as a parameterized multimodal learning task that grounds report text to topology/attributes, reconstructs motion, and refines interactions. Performance is reported as outperformance versus baselines on reconstruction fidelity metrics (accident point accuracy, collision consistency) within that dataset. This constitutes standard supervised learning with external ground-truth measurements from the source investigations; the 'quantitatively verifiable' claim rests on comparison to held-out real data rather than reducing to a self-fit, self-definition, or self-citation chain. No equations, ansatzes, or uniqueness theorems are invoked that collapse to the inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that report text supplies enough information for physical grounding and that the learning process produces verifiable reconstructions rather than fitted artifacts; no independent physical simulation engine or external validation is mentioned.

free parameters (1)

multimodal model parameters
The parameterized learning problem implies numerous weights and hyperparameters fitted to the CISS-REC dataset to achieve the reported fidelity.

axioms (1)

domain assumption Public textual reports contain sufficient semantic information to ground to road topology and participant attributes
Invoked in the formulation of accident reconstruction as a parameterized multimodal learning problem.

pith-pipeline@v0.9.0 · 5455 in / 1330 out tokens · 95495 ms · 2026-05-09T19:55:08.355101+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

59 extracted references · 5 canonical work pages · 2 internal anchors

[1]

D. Lord, F. Mannering, The statistical analysis of crash-frequency data:Areviewandassessmentofmethodologicalalternatives, Trans- portation research part A: policy and practice 44 (2010) 291–305

2010
[2]

P. T. Savolainen, F. L. Mannering, D. Lord, M. A. Quddus, The statistical analysis of highway crash-injury severities: A review and assessment of methodological alternatives, Accident Analysis & Prevention 43 (2011) 1666–1676

2011
[3]

Arteaga, A

C. Arteaga, A. Paz, J. Park, Injury severity on traffic crashes: A text mining with an interpretable machine-learning approach, Safety Science 132 (2020) 104988. Yanchen Guan et al.:Preprint submitted to Elsevier Page 20 of 22

2020
[4]

S.Lee,R.Arvin,A.J.Khattak, Advancinginvestigationofautomated vehicle crashes using text analytics of crash narratives and bayesian analysis, Accident Analysis & Prevention 181 (2023) 106932

2023
[5]

Tiwari, D

G. Tiwari, D. Mohan, G. Agrawal, Transport planning and traffic safety: making cities, roads, and vehicles safer, CRC Press, 2018

2018
[6]

L.Hu,X.Bao,M.Lin,C.Yu,F.Wang, Researchonriskydrivingbe- havior evaluation model based on cidas real data, Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering 235 (2021) 2176–2187

2021
[7]

D. Otte, M. Jänsch, C. Haasper, Injury protection and accident causation parameters for vulnerable road users based on german in- depthaccidentstudygidas,AccidentAnalysis&Prevention44(2012) 149–153

2012
[8]

Kreiss, G

J.-P. Kreiss, G. Feng, J. Krampe, M. Meyer, T. Niebuhr, C. Pastor, J. Dobberstein, Extrapolation of gidas accident data to europe, in: Proceedings of the 24th ESV Conference Proceedings, NHTSA Washington, DC, 2015

2015
[9]

Imbriani,M.A.Romero, Crashdatabasesinaustralasia,theeuropean union, and the united states: review and prospects for improvement, Transportation research record 2386 (2013) 128–136

A.Montella,D.Andreassen,A.P.Tarko,S.Turner,F.Mauriello,L.L. Imbriani,M.A.Romero, Crashdatabasesinaustralasia,theeuropean union, and the united states: review and prospects for improvement, Transportation research record 2386 (2013) 128–136

2013
[10]

Zhang, Q

Z. Zhang, Q. He, J. Gao, M. Ni, A deep learning approach for detecting traffic accidents from social media data, Transportation research part C: emerging technologies 86 (2018) 580–596

2018
[11]

C. S. Shin, W. Pang, C. Li, F. Bai, F. Ahmad, J. Paek, R. Govindan, Recap: 3d traffic reconstruction, in: Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, 2024, pp. 1252–1267

2024
[12]

J. Beck, R. Arvin, S. Lee, A. Khattak, S. Chakraborty, Automated vehicle data pipeline for accident reconstruction: New insights from lidar, camera, and radar data, Accident Analysis & Prevention 180 (2023) 106923

2023
[13]

Wach, Calculation reliability in vehicle accident reconstruction, Forensic science international 263 (2016) 27–38

W. Wach, Calculation reliability in vehicle accident reconstruction, Forensic science international 263 (2016) 27–38

2016
[14]

A.Guo,Y.Zhou,H.Tian,C.Fang,Y.Sun,W.Sun,X.Gao,A.T.Luu, Y. Liu, Z. Chen, Sovar: Build generalizable scenarios from accident reports for autonomous driving testing, in: Proceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering, 2024, pp. 268–280

2024
[15]

F. Tan, S. Feng, V. Ordonez, Text2scene: Generating compositional scenes from textual descriptions, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 6710–6719

2019
[16]

Z.Li,Z.Cui,H.Liao,J.Ash,G.Zhang,C.Xu,Y.Wang, Steeringthe future: Redefining intelligent transportation systems with foundation models, Chain 1 (2024) 46–53

2024
[17]

R.L.Griffin,S.Carroll,J.O.Jansen, Automaticcollisionnotification availability and emergency response times following vehicle colli- sion—an analysis of the 2017 crash investigation sampling system, Traffic injury prevention 21 (2020) S135–S139

2017
[18]

Z.Zheng,Z.Wang,L.Zhu,H.Jiang, Determinantsofthecongestion causedbyatrafficaccidentinurbanroadnetworks, AccidentAnalysis & Prevention 136 (2020) 105327

2020
[19]

D. E. Struble, J. D. Struble, Automotive accident reconstruction: practices and principles, CRC Press, 2020

2020
[20]

F. A. Fernandes, R. J. Alves de Sousa, M. Ptak, Application of numerical methods for accident reconstruction and forensic analysis, in:HeadInjurySimulationinRoadTrafficAccidents,Springer,2018, pp. 59–98

2018
[21]

D.S.Ryan,E.Zarei, Accidentinvestigationprocessesandtechniques insociotechnicalsystems, in:SafetyCausationAnalysisinSociotech- nicalSystems:AdvancedModelsandTechniques,Springer,2024,pp. 21–45

2024
[22]

R. W. Rivers, Technical Traffic Crash Investigators’ Handbook:(level 3):aTechnicalReference,Training,InvestigationandReconstruction Manual, Charles C Thomas Publisher, 2010

2010
[23]

R. W. Rivers, Evidence in traffic crash investigation and recon- struction: identification, interpretation and analysis of evidence, and the traffic crash investigation and reconstruction process, Charles C Thomas Publisher, 2006

2006
[24]

S. I. Mohammed, An overview of traffic accident investigation using different techniques, Automotive experiences 6 (2023) 68–79

2023
[25]

G. Vida, G. Melegh, Á. Süveges, N. Wenszky, Á. Török, Analysis of uav flight patterns for road accident site investigation, Vehicles 5 (2023) 1707–1726

2023
[26]

S. Su, W. Liu, K. Li, G. Yang, C. Feng, J. Ming, G. Liu, S. Liu, Z.Yin, Developinganunmannedaerialvehicle-basedrapidmapping systemfortrafficaccidentinvestigation, Australianjournalofforensic sciences 48 (2016) 454–468

2016
[27]

Jiang, W

S. Jiang, W. Jiang, L. Wang, Unmanned aerial vehicle-based pho- togrammetric 3d mapping: A survey of techniques, applications, and challenges, IEEE Geoscience and Remote Sensing Magazine 10 (2021) 135–171

2021
[28]

Lemmens, Terrestrial laser scanning, in: Geo-information: technologies, applications and the environment, Springer, 2011, pp

M. Lemmens, Terrestrial laser scanning, in: Geo-information: technologies, applications and the environment, Springer, 2011, pp. 101–121

2011
[29]

Scherer, J

M. Scherer, J. L. Lerma, From the conventional total station to the prospective image assisted photogrammetric scanning total station: Comprehensivereview, JournalofSurveyingEngineering135(2009) 173–178

2009
[30]

Clamann, A

M. Clamann, A. J. Khattak, K. Clark, et al., Advancing crash investi- gationwithconnectedandautomatedvehicle data,TechnicalReport, Collaborative Sciences Center for Road Safety, 2021

2021
[31]

Dhanam, P

B. Dhanam, P. Marichamy, B. Mohan, E. Manoj, J. Johnson, K. Bal- avignesh, Event data recorder for investigation, legal claim and fault analysis in vehicles, in: 2025 3rd International Conference on Artificial Intelligence and Machine Learning Applications Theme: Healthcare and Internet of Things (AIMLA), IEEE, 2025, pp. 1–5

2025
[32]

K. L. Carper, Forensic engineering, CRC press, 2000

2000
[33]

Raviv, B

G. Raviv, B. Fishbain, A. Shapira, Analyzing risk factors in crane- relatednear-missandaccidentreports, Safetyscience91(2017)192– 205

2017
[34]

Faizan, A

K. Faizan, A. Abid, Forensic investigation of road traffic accident cases in pakistan and types of physical evidence, Pakistan Social Sciences Review 5 (2021) 405–422

2021
[35]

H. W. Smith, Physical evidence in the investigation of traffic acci- dents, J. Crim. L. Criminology & Police Sci. 48 (1957) 93

1957
[36]

D. Liu, D. Li, N. Sze, H. Ding, Y. Song, An integrated data-and theory-driven crash severity model, Accident Analysis & Prevention 193 (2023) 107282

2023
[37]

A.Gadotti,L.Rocher,F.Houssiau,A.-M.Creţu,Y.-A.DeMontjoye, Anonymization:Theimperfectscienceofusingdatawhilepreserving privacy, Science advances 10 (2024) eadn7053

2024
[38]

K. N. Morehouse, B. Kurdi, B. A. Nosek, Responsible data sharing: Identifyingandremedyingpossiblere-identificationofhumanpartic- ipants., American Psychologist (2024)

2024
[39]

M. M. Hossain, H. Zhou, S. Das, Data mining approach to explore emergency vehicle crash patterns: A comparative study of crash severityinemergencyandnon-emergencyresponsemodes, Accident Analysis & Prevention 191 (2023) 107217

2023
[40]

la importancia del atestado oficial

J. Font Mezquita, Reconstrucción de accidentes:“la importancia del atestado oficial”, Securitas Vialis 4 (2012) 9–15

2012
[41]

R. R. Rider, The impact of new technology on crash reconstruction, Tarleton State University, 2017

2017
[42]

Ball, Working with images in daily life and police practice: an assessment of the documentary tradition, Qualitative Research 5 (2005) 499–521

M. Ball, Working with images in daily life and police practice: an assessment of the documentary tradition, Qualitative Research 5 (2005) 499–521

2005
[43]

M. L. Komter, From talk to text: The interactional construction of a policerecord, ResearchonLanguageandSocialinteraction39(2006) 201–228

2006
[44]

Y. Chen, Q. Zhang, F. Yu, Transforming traffic accident investiga- tions:avirtual-real-fusionframeworkforintelligent3dtrafficaccident reconstruction, Complex & Intelligent Systems 11 (2025) 76

2025
[45]

P. Jiao, Q. Miao, M. Zhang, W. Zhao, A virtual reality method for digitally reconstructing traffic accidents from videos or still images, Forensic science international 292 (2018) 176–180. Yanchen Guan et al.:Preprint submitted to Elsevier Page 21 of 22

2018
[46]

C. Li, K. Zhou, T. Liu, Y. Wang, M. Zhuang, H.-a. Gao, B. Jin, H. Zhao, Avd2: Accident video diffusion for accident video descrip- tion, arXiv preprint arXiv:2502.14801 (2025)

work page arXiv 2025
[47]

AccidentSim: Generating Vehicle Collision Videos with Physically Realistic Collision Trajectories from Real-World Accident Reports

X. Zhang, Q. Zhang, L. Han, Q. Qu, X. Chen, Accidentsim: Gen- erating physically realistic vehicle collision videos from real-world accident reports, arXiv preprint arXiv:2503.20654 (2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[48]

M. Li, W. Ding, H. Lin, Y. Lyu, Y. Yao, Y. Zhang, D. Zhao, Crasha- gent: Crash scenario generation via multi-modal reasoning, arXiv preprint arXiv:2505.18341 (2025)

work page arXiv 2025
[49]

Bertolazzi, M

E. Bertolazzi, M. Frego, On the g2 hermite interpolation problem with clothoids, Journal of Computational and Applied Mathematics 341 (2018) 99–116

2018
[50]

Krajzewicz, J

D. Krajzewicz, J. Erdmann, M. Behrisch, L. Bieker, et al., Recent development and applications of sumo-simulation of urban mobility, International journal on advances in systems and measurements 5 (2012) 128–138

2012
[51]

World Simulation with Video Foundation Models for Physical AI

A.Ali,J.Bai,M.Bala,Y.Balaji,A.Blakeman,T.Cai,J.Cao,T.Cao, E. Cha, Y.-W. Chao, et al., World simulation with video foundation models for physical ai, arXiv preprint arXiv:2511.00062 (2025)

work page internal anchor Pith review arXiv 2025
[52]

H. Han, M. Zhang, M. Hou, F. Zhang, Z. Wang, E. Chen, H. Wang, J.Ma,Q.Liu, Stgcn:aspatial-temporalawaregraphlearningmethod for poi recommendation, in: 2020 IEEE International Conference on Data Mining (ICDM), IEEE, 2020, pp. 1052–1057

2020
[53]

Schumaker, Spline functions: basic theory, Cambridge university press, 2007

L. Schumaker, Spline functions: basic theory, Cambridge university press, 2007

2007
[54]

G.VanHoudt,C.Mosquera,G.Nápoles, Areviewonthelongshort- term memory model, Artificial intelligence review 53 (2020) 5929– 5955

2020
[55]

Wayformer: Motion forecasting via simple & efficient attention networks.arXiv preprint arXiv:2207.05844, 2022

N. Nayakanti, R. Al-Rfou, A. Zhou, K. Goel, K. S. Refaat, B. Sapp, Wayformer: Motion forecasting via simple & efficient attention net- works, arXiv preprint arXiv:2207.05844 (2022)

work page arXiv 2022
[56]

Z. Zhou, L. Ye, J. Wang, K. Wu, K. Lu, Hivt: Hierarchical vector transformerformulti-agentmotionprediction, in:Proceedingsofthe IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8823–8833

2022
[57]

J. Gao, C. Sun, H. Zhao, Y. Shen, D. Anguelov, C. Li, C. Schmid, Vectornet: Encoding hd maps and agent dynamics from vectorized representation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11525–11533

2020
[58]

Steffan, A

H. Steffan, A. Moser, The collision and trajectory models of PC- CRASH, Technical Report, SAE Technical Paper, 1996

1996
[59]

1128–1135

X.Zhang,X.Jin,J.Shen, Virtualreconstructionoftwotypesoftraffic accident by the tire marks, in: International Conference on Artificial Reality and Telexistence, Springer, 2006, pp. 1128–1135. Yanchen Guan et al.:Preprint submitted to Elsevier Page 22 of 22

2006