pith. machine review for the scientific record. sign in

arxiv: 2604.17841 · v1 · submitted 2026-04-20 · 💻 cs.RO

Recognition: unknown

Driving risk emerges from the required two-dimensional joint evasive acceleration

Hao Cheng , Yanbo Jiang , Wenhao Yu , Rui Zhou , Jiang Bian , Keyu Chen , Zhiyuan Liu , Heye Huang
show 4 more authors
Hailun Zhang Fang Zhang Jianqiang Wang Sifa Zheng
Authors on Pith no claims yet

Pith reviewed 2026-05-10 04:56 UTC · model grok-4.3

classification 💻 cs.RO
keywords driving riskevasive accelerationautonomous drivingcollision avoidancetime-to-collisionrisk quantificationvehicle interactionssafety benchmarks
0
0 comments X

The pith

Driving risk is the minimum constant relative acceleration needed in any direction to prevent a collision.

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

Standard benchmarks measure risk with time-to-collision, which treats the problem as one-dimensional closing speed. This paper defines evasive acceleration as the smallest constant relative acceleration vector that, applied in the best direction, makes the paths miss each other. On five public datasets plus more than 600 real crashes, the measure triggers the earliest statistically significant warnings at every threshold, separates collision outcomes more cleanly, and retains 54.2-241.4 percent more predictive information than baselines. It adds substantial non-redundant content even when combined with existing metrics. Because the definition follows directly from relative motion without extra parameters, it supplies a physically grounded alternative for safety assessment.

Core claim

Evasive acceleration defines risk as the minimum magnitude of a constant relative acceleration vector, evaluated over all possible directions, that is required to alter the relative motion and ensure the interaction remains collision-free. This two-dimensional formulation captures the joint evasive actions of both vehicles. When applied to interaction data from five open datasets and over 600 real crashes, percentile thresholds derived from it trigger the earliest statistically significant warnings, achieve superior discrimination between collision and non-collision outcomes, and retain 54.2-241.4 percent more information than baseline methods.

What carries the argument

Evasive acceleration: the minimum magnitude of a constant relative acceleration vector required to achieve collision-free motion, obtained by testing all possible avoidance directions.

If this is right

  • EA provides the earliest statistically significant warning across all tested thresholds.
  • EA achieves the best discrimination between eventual collision and non-collision outcomes.
  • Information retention improves by 54.2-241.4 percent over all compared baselines.
  • Adding EA to existing methods produces 17.5-95.5 times more information gain than adding existing methods to EA.
  • EA captures most outcome-relevant information already present in baselines while contributing substantial nonredundant content.

Where Pith is reading between the lines

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

  • Vehicle planners could optimize trajectories by minimizing evasive acceleration directly rather than by satisfying separate time-to-collision thresholds.
  • In dense multi-agent scenes the maximum pairwise EA could serve as a single scalar risk signal for priority decisions.
  • The constant-acceleration model could be relaxed to linear or piecewise acceleration profiles to test whether risk estimates improve in high-curvature or braking scenarios.

Load-bearing premise

That the smallest constant relative acceleration sufficient to clear a collision in the optimal direction accurately represents real-world risk where accelerations vary over time and multiple agents interact.

What would settle it

In a large set of real crashes, EA-based warnings do not occur statistically earlier or discriminate outcomes better than time-to-collision warnings at matched thresholds.

Figures

Figures reproduced from arXiv: 2604.17841 by Fang Zhang, Hailun Zhang, Hao Cheng, Heye Huang, Jiang Bian, Jianqiang Wang, Keyu Chen, Rui Zhou, Sifa Zheng, Wenhao Yu, Yanbo Jiang, Zhiyuan Liu.

Figure 1
Figure 1. Figure 1: Evasive acceleration (EA) as a new paradigm for risk quantification: motivation, concept and validation. a, Illustration of a traffic scene evaluated by TTC, its variants, and existing risk quantification methods. TTC and its derivatives have long shaped safety assessment across training, testing, and regulatory settings. b, Conceptual illustration of the core limitation of the TTC paradigm. i, TTC evaluat… view at source ↗
Figure 2
Figure 2. Figure 2: Representative interaction cases showing how EA captures risk evolution and conflict resolution. a, Merging interaction at an intersection. Panels i and iii show two representative instants of the interaction, with trajectory colours indicating EA values. Panels ii and iv show the corresponding collision-avoiding and non-collision-avoiding sets in two￾dimensional joint-acceleration space. The red solid arr… view at source ↗
Figure 3
Figure 3. Figure 3: Statistical separability of precrash states and potential conflicts. a, Precision–recall (PR) and receiver operating characteristic (ROC) curves for four precrash lead windows: [−0.5,−0.1], [−1.0,−0.1], [−1.5,−0.1], and [−2.0,−0.1] s. EA shows the strongest overall discrimination across the four lead windows, with its advantage becoming more pronounced at earlier precrash horizons. b, Summary discriminatio… view at source ↗
Figure 4
Figure 4. Figure 4: Early warning performance under percentile-aligned false-alarm constraints. a, Definition of warning lead time (WLT) under the sustained warning criterion. i, If threshold exceedance persists to the last effective precollision frame, WLT is measured from the first such exceedance. ii, If the warning disappears and later reappears, only the final sustained exceedance is counted. iii, If no sustained warning… view at source ↗
Figure 5
Figure 5. Figure 5: EA exhibits the best standalone discrimination, the lowest residual uncertainty, the greatest retained information, and a pronounced directional asymmetry in incremental information relative to existing baselines. All methods are evaluated on identical interaction states at each lead time. Real crash cases are aligned by the true impact time (t = 0), whereas noncrash potential-conflict episodes are aligned… view at source ↗
read the original abstract

Most autonomous driving safety benchmarks use time-to-collision (TTC) to assess risk and guide safe behaviour. However, TTC-based methods treat risk as a one-dimensional closing problem, despite the inherently two-dimensional nature of collision avoidance, and therefore cannot faithfully capture risk or its evolution over time. Here, we report evasive acceleration (EA), a hyperparameter-free and physically interpretable two-dimensional paradigm for risk quantification. By evaluating all possible directions of collision avoidance, EA defines risk as the minimum magnitude of a constant relative acceleration vector required to alter the relative motion and make the interaction collision-free. Using interaction data from five open datasets and more than 600 real crashes, we derive percentile-based warning thresholds and show that EA provides the earliest statistically significant warning across all thresholds. Moreover, EA provides the best discrimination of eventual collision outcomes and improves information retention by 54.2-241.4% over all compared baselines. Adding EA to existing methods yields 17.5-95.5 times more information gain than adding existing methods to EA, indicating that EA captures much of the outcome-relevant information in existing methods while contributing substantial additional nonredundant information. Overall, EA better captures the structure of collision risk and provides a foundation for next-generation autonomous driving systems.

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.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces evasive acceleration (EA) as a two-dimensional risk metric for driving interactions. EA is defined as the smallest magnitude of a constant relative acceleration vector, applied in an optimal direction, that ensures no collision occurs between agents over the time horizon. The authors evaluate EA using data from five open datasets and more than 600 real-world crashes, deriving percentile-based warning thresholds. They claim that EA provides the earliest statistically significant warnings, best discriminates collision outcomes, improves information retention by 54.2-241.4% over baselines, and adds substantial non-redundant information when combined with existing methods.

Significance. If the central modeling assumptions hold, this work provides a physically grounded alternative to one-dimensional metrics like TTC, with empirical support from real crash data demonstrating improved predictive power and information content. The use of multiple datasets and crash records strengthens the empirical claims. The hyperparameter-free construction and direct use of relative motion principles are notable strengths.

major comments (2)
  1. [§3] §3 (EA definition): EA is constructed as the minimum scalar magnitude of a constant relative acceleration vector that produces a collision-free trajectory. This modeling choice implicitly assumes that a single fixed acceleration suffices over the full horizon; it does not address cases where optimal evasion requires time-varying acceleration (e.g., initial braking then steering) or responses to additional agents. Because the percentile thresholds, discrimination results, and 54.2–241.4% information-retention gains rest directly on this scalar proxy, the reported superiority over baselines could be sensitive to the constant-acceleration restriction.
  2. [Results] Results section (information-gain and threshold analysis): The claims that EA yields 17.5–95.5× more incremental information gain and provides the earliest statistically significant warning across thresholds require explicit reporting of the exact statistical procedure (e.g., which test, degrees of freedom, multiple-comparison correction) and the per-dataset sample sizes at each percentile. Without these, it is difficult to assess whether the numerical improvements are robust or partly artifacts of the constant-acceleration modeling choice.
minor comments (2)
  1. [Abstract] Abstract and §3: The phrase 'hyperparameter-free' is used, yet practical computation of EA necessarily involves discretization of direction and time; these choices should be stated explicitly as implementation parameters even if they do not appear in the final scalar value.
  2. [§3] Notation: The relative acceleration vector and the occupied-region intersection test are central; a compact equation or pseudocode block early in §3 would improve readability for readers unfamiliar with the exact geometric formulation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which have helped us clarify key aspects of the work. We address each major comment below and have revised the manuscript to strengthen the presentation of our modeling choices and statistical reporting.

read point-by-point responses

  1. Referee: [§3] §3 (EA definition): EA is constructed as the minimum scalar magnitude of a constant relative acceleration vector that produces a collision-free trajectory. This modeling choice implicitly assumes that a single fixed acceleration suffices over the full horizon; it does not address cases where optimal evasion requires time-varying acceleration (e.g., initial braking then steering) or responses to additional agents. Because the percentile thresholds, discrimination results, and 54.2–241.4% information-retention gains rest directly on this scalar proxy, the reported superiority over baselines could be sensitive to the constant-acceleration restriction.

    Authors: The constant-acceleration formulation is a deliberate modeling choice that yields a hyperparameter-free, physically interpretable metric representing the minimum fixed relative acceleration needed to ensure collision-free trajectories. This aligns with the paper's emphasis on a simple, two-dimensional risk measure grounded in relative motion. While we acknowledge that time-varying accelerations or multi-agent responses may be optimal in some scenarios, the empirical evaluation across five open datasets and over 600 real crashes shows consistent superiority in warning timing, discrimination, and information retention. To address the concern directly, we have added a dedicated limitations paragraph in §3 explaining the rationale for the constant-acceleration assumption, noting its conservative nature for short-horizon interactions, and outlining potential extensions to time-varying or multi-agent cases. This is a partial revision. revision: partial

  2. Referee: [Results] Results section (information-gain and threshold analysis): The claims that EA yields 17.5–95.5× more incremental information gain and provides the earliest statistically significant warning across thresholds require explicit reporting of the exact statistical procedure (e.g., which test, degrees of freedom, multiple-comparison correction) and the per-dataset sample sizes at each percentile. Without these, it is difficult to assess whether the numerical improvements are robust or partly artifacts of the constant-acceleration modeling choice.

    Authors: We agree that full transparency in statistical methods is necessary for assessing robustness. In the revised manuscript, we have expanded the Results section to explicitly describe the statistical procedures (paired Wilcoxon signed-rank tests with Bonferroni correction for multiple thresholds, with degrees of freedom and p-value thresholds reported), and we now include per-dataset sample sizes at each percentile level for all analyses. These details confirm that the reported gains in information retention and earlier warnings are consistent across independent datasets and not artifacts of the modeling choice. This is a full revision. revision: yes

Circularity Check

0 steps flagged

No circularity: EA defined from kinematics and validated empirically on external data

full rationale

The paper defines EA directly as the minimum-magnitude constant relative acceleration vector that renders an interaction collision-free, obtained by exhaustive search over directions in the two-dimensional relative-motion plane. This construction follows from standard relative kinematics and does not presuppose any fitted parameter or outcome label that is later re-predicted. Percentile thresholds are computed once from the empirical distribution on five open datasets; discrimination, information-retention, and incremental-gain statistics are then obtained by applying those fixed thresholds to held-out crash records and comparing against independent baselines. No equation reduces to its own input by construction, no self-citation supplies a load-bearing uniqueness theorem, and no ansatz is smuggled in. The modeling choice of constant acceleration is an explicit assumption whose fidelity can be tested externally; it does not create a definitional loop.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the physical definition of EA and empirical comparisons; no free parameters are introduced as the method is stated to be hyperparameter-free.

axioms (1)
  • domain assumption Collision risk can be quantified by the minimum constant relative acceleration vector that alters motion to collision-free state.
    Invoked in the definition of EA as the core modeling choice for 2D avoidance.
invented entities (1)
  • Evasive acceleration (EA) no independent evidence
    purpose: Quantify 2D collision risk
    Newly defined quantity introduced to replace or augment 1D TTC.

pith-pipeline@v0.9.0 · 5556 in / 1206 out tokens · 39963 ms · 2026-05-10T04:56:07.747841+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

69 extracted references · 14 canonical work pages · 2 internal anchors

  1. [1]

    Feng, S.et al.Breaking through safety performance stagnation in autonomous vehicles with dense learning. Nat. Commun.(2026)

    2026

  2. [2]

    & Abbink, D

    Kolekar, S., De Winter, J. & Abbink, D. Human-like driving behaviour emerges from a risk-based driver model. Nat. communications11, 4850 (2020)

    2020

  3. [3]

    Feng, S.et al.Dense reinforcement learning for safety validation of autonomous vehicles.Nature615, 620–627 (2023)

    2023

  4. [4]

    & Guo, F

    Qian, C., Xu, J., Xing, X. & Guo, F. Test case sampling optimization for safety validation of automated driving systems.Nat. Commun.(2026)

    2026

  5. [5]

    & Chitta, K

    Dauner, D., Hallgarten, M., Geiger, A. & Chitta, K. Parting with misconceptions about learning-based vehicle motion planning. InConference on Robot Learning, 1268–1281 (PMLR, 2023)

    2023

  6. [6]

    Neural Inf

    Dauner, D.et al.Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking.Adv. Neural Inf. Process. Syst.37, 28706–28719 (2024)

    2024

  7. [7]

    InConference on Robot Learning, 4709–4722 (PMLR, 2025)

    Cao, W.et al.Pseudo-simulation for autonomous driving. InConference on Robot Learning, 4709–4722 (PMLR, 2025)

    2025

  8. [8]

    X.et al.Behavioral safety assessment towards large-scale deployment of autonomous vehicles.arXiv preprint arXiv:2505.16214(2025)

    Liu, H. X.et al.Behavioral safety assessment towards large-scale deployment of autonomous vehicles.arXiv preprint arXiv:2505.16214(2025)

    work page arXiv 2025

  9. [9]

    Cui, Y .et al.Vp-autotest: A virtual-physical fusion autonomous driving testing platform.arXiv preprint arXiv:2512.07507(2025)

    work page arXiv 2025

  10. [10]

    & Wang, S

    Fan, Y ., Li, Y . & Wang, S. Risk-aware self-consistent imitation learning for trajectory planning in autonomous driving. InEuropean Conference on Computer Vision, 270–287 (Springer, 2024)

    2024

  11. [11]

    Guo, K., Liu, H., Wu, X., Pan, J. & Lv, C. ipad: Iterative proposal-centric end-to-end autonomous driving. arXiv preprint arXiv:2505.15111(2025)

    work page arXiv 2025

  12. [12]

    Li, Z.et al.Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation.arXiv preprint arXiv:2406.06978(2024)

    work page internal anchor Pith review arXiv 2024

  13. [13]

    Yao, W.et al.Drivesuprim: Towards precise trajectory selection for end-to-end planning.arXiv preprint arXiv:2506.06659(2025)

    work page arXiv 2025

  14. [14]

    Li, Z.et al.Generalized trajectory scoring for end-to-end multimodal planning.arXiv preprint arXiv:2506.06664 (2025)

    work page arXiv 2025

  15. [15]

    Diver: Reinforced diffusion breaks imitation bottlenecks in end-to-end autonomous driving.arXiv preprint arXiv:2507.04049, 2025

    Song, Z.et al.Diver: Reinforced diffusion breaks imitation bottlenecks in end-to-end autonomous driving. arXiv preprint arXiv:2507.04049(2025)

    work page internal anchor Pith review arXiv 2025

  16. [16]

    B., Loo, J

    Reddy, R., Barua, H. B., Loo, J. Y ., Nguyen, T. T. & Krishnasamy, G. Rapid: Real-time deterministic trajectory planning via diffusion behavior priors for safe and efficient autonomous driving.arXiv preprint arXiv:2602.07339(2026). 17.Yang, Z.et al.Read: End-to-end autonomous driving made safer with efficient reinforcement learning (2025). Preprint

    work page arXiv 2026

  17. [17]

    Plannerrft: Reinforcing diffu- sion planners through closed-loop and sample-efficient fine- tuning.arXiv preprint arXiv:2601.12901, 2026

    Li, H.et al.Plannerrft: Reinforcing diffusion planners through closed-loop and sample-efficient fine-tuning. arXiv preprint arXiv:2601.12901(2026)

    work page arXiv 2026

  18. [18]

    Song, Y .et al.Senna-2: Aligning vlm and end-to-end driving policy for consistent decision making and planning.arXiv preprint arXiv:2603.11219(2026)

    work page arXiv 2026

  19. [19]

    Nolte, M.et al.A review of conceptualizations of safety and risk in current automated driving regulation.arXiv preprint arXiv:2502.06594(2025). 19/23

    work page arXiv 2025

  20. [20]

    & Den Camp, O

    de Gelder, E. & Den Camp, O. O. Scenario-based assessment of automated driving systems: How (not) to parameterize scenarios? In2024 IEEE International Automated Vehicle Validation Conference (IAVVC), 1–7 (IEEE, 2024)

    2024

  21. [21]

    & Bärgman, J

    Olleja, P., Markkula, G. & Bärgman, J. Validation of human benchmark models for automated driving system approval: How competent and careful are they really?Accid. Analysis & Prev.213, 107922 (2025)

    2025

  22. [22]

    & Dozza, M

    Rasch, A., Panero, G., Boda, C.-N. & Dozza, M. How do drivers overtake pedestrians? evidence from field test and naturalistic driving data.Accid. Analysis & Prev.139, 105494 (2020)

    2020

  23. [23]

    E.et al.Indicators of driver adaptation to forward collision warnings: A naturalistic driving evaluation (2019)

    Nodine, E. E.et al.Indicators of driver adaptation to forward collision warnings: A naturalistic driving evaluation (2019). Source bibliography lacked journal information

    2019

  24. [24]

    B., den Camp Olaf, O.et al.Scenario metrics for the safety assurance framework of automated vehicles: A review of its application.Vehicles7, 100 (2025)

    Singh, T., Fouad, H.-S., Sergi, V . B., den Camp Olaf, O.et al.Scenario metrics for the safety assurance framework of automated vehicles: A review of its application.Vehicles7, 100 (2025)

    2025

  25. [25]

    Venthuruthiyil, S. P. & Chunchu, M. Anticipated collision time (act): A two-dimensional surrogate safety indicator for trajectory-based proactive safety assessment.Transp. research part C: emerging technologies139, 103655 (2022)

    2022

  26. [26]

    & Keyvan-Ekbatani, M

    Guo, H., Xie, K. & Keyvan-Ekbatani, M. Modeling driver’s evasive behavior during safety–critical lane changes: Two-dimensional time-to-collision and deep reinforcement learning.Accid. Analysis & Prev.186, 107063 (2023)

    2023

  27. [27]

    Mathematics12, 1905 (2024)

    Fu, X.et al.Research on vehicle aeb control strategy based on safety time–safety distance fusion algorithm. Mathematics12, 1905 (2024)

    1905

  28. [28]

    & Chen, K

    Wang, Y ., Li, Z., Liu, P., Xu, C. & Chen, K. Surrogate safety measures for traffic oscillations based on empirical vehicle trajectories prior to crashes.Transp. research part C: emerging technologies161, 104543 (2024)

    2024

  29. [29]

    & Liu, Y .-C

    Jafari, A. & Liu, Y .-C. Pedestrians’ safety using projected time-to-collision to electric scooters.Nat. communi- cations15, 5701 (2024)

    2024

  30. [30]

    & Aramrattana, M

    Behera, A., Kharrazi, S., Frisk, E. & Aramrattana, M. An improved 2d time-to-collision for articulated vehicles: predicting sideswipe and rear-end collisions.arXiv e-printsarXiv–2507 (2025)

    2025

  31. [31]

    & Yeo, H

    Lin, T., Jin, Z., Choi, S. & Yeo, H. Predicted lane-based time to collision: A novel surrogate safety measure for pedestrian potential risk evaluation at non-signalized intersections.Transp. Res. Interdiscip. Perspectives34, 101705 (2025)

    2025

  32. [32]

    Tang, F., Liu, Z., Wang, Z. & Li, N. Analysis of traffic conflict characteristics and key factors influencing severity in expressway interchange diverging areas: Insights from a chinese freeway safety study.Sustainability 17, 8419 (2025)

    2025

  33. [33]

    J., Flannagan, C

    Yan, X., Feng, S., LeBlanc, D. J., Flannagan, C. & Liu, H. X. Evaluation of automated driving system safety metrics with logged vehicle trajectory data.IEEE Transactions on Intell. Transp. Syst.25, 8913–8925 (2024)

    2024

  34. [34]

    & Kim, S

    Choi, J. & Kim, S. Predictive risk-aware reinforcement learning for autonomous vehicles using safety potential. Electronics14, 4446 (2025)

    2025

  35. [35]

    & Rosman, G

    Biswas, A., Gideon, J., Tamura, K. & Rosman, G. Modeling drivers’ risk perception via attention to improve driving assistance.arXiv preprint arXiv:2409.04738(2024)

    work page arXiv 2024

  36. [36]

    Ai2-active safety: Ai-enabled interaction-aware active safety analysis with vehicle dynamics.arXiv preprint arXiv:2505.00322, 2025

    Wu, K.et al.Ai2-active safety: Ai-enabled interaction-aware active safety analysis with vehicle dynamics. arXiv preprint arXiv:2505.00322(2025)

    work page arXiv 2025

  37. [37]

    & Yang, S

    Lu, D., Du, H., Wu, Z. & Yang, S. Risk assessment in autonomous driving: a comprehensive survey of risk sources, methodologies, and system architectures.Auton. Intell. Syst.5, 24 (2025)

    2025

  38. [38]

    Diwakar, P., Landge, V . S. & Jain, U. Evaluating the relationship between surrogate safety measures and traffic event severity in terms of human perception of danger: A perspective under indian traffic conditions.Appl. Sci. 13, 12100 (2023). 20/23

    2023

  39. [39]

    & Miranda-Moreno, L

    Do, W., Saunier, N. & Miranda-Moreno, L. Evaluation of conventional surrogate indicators of safety for connected and automated vehicles in car following at signalized intersections.Transp. Res. Rec.2679, 1118– 1133 (2025)

    2025

  40. [40]

    & Abdulhai, B

    ElSamadisy, O., Shi, T., Smirnov, I. & Abdulhai, B. Safe, efficient, and comfortable reinforcement-learning- based car-following for avs with an analytic safety guarantee and dynamic target speed.Transp. research record 2678, 643–661 (2024)

    2024

  41. [41]

    Westhofen, L.et al.Criticality metrics for automated driving: A review and suitability analysis of the state of the art.Arch. Comput. Methods Eng.30, 1–35 (2023)

    2023

  42. [42]

    Analysis & Prev

    Zhou, R.et al.How would autonomous vehicles behave in real-world crash scenarios?Accid. Analysis & Prev. 202, 107572 (2024)

    2024

  43. [43]

    & Bian, J

    Zhou, R., Huang, H., Zhang, G., Zhou, H. & Bian, J. Crash-based safety testing of autonomous vehicles: Insights from generating safety-critical scenarios based on in-depth crash data.IEEE Transactions on Intell. Transp. Syst.(2025)

    2025

  44. [44]

    & Meng, X

    Zheng, L., Ismail, K. & Meng, X. Traffic conflict techniques for road safety analysis: open questions and some insights.Can. journal civil engineering41, 633–641 (2014)

    2014

  45. [45]

    & De Ceunynck, T

    Johnsson, C., Laureshyn, A. & De Ceunynck, T. In search of surrogate safety indicators for vulnerable road users: a review of surrogate safety indicators.Transp. reviews38, 765–785 (2018)

    2018

  46. [46]

    R., Rao, K

    Sarkar, D. R., Rao, K. R. & Chatterjee, N. A review of surrogate safety measures on road safety at unsignalized intersections in developing countries.Accid. Analysis & Prev.195, 107380 (2024)

    2024

  47. [47]

    & Eckstein, L

    Krajewski, R., Bock, J., Kloeker, L. & Eckstein, L. The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems. In2018 21st international conference on intelligent transportation systems (ITSC), 2118–2125 (IEEE, 2018)

    2018

  48. [48]

    In2022 IEEE Intelligent Vehicles Symposium (IV), 958–964 (IEEE, 2022)

    Moers, T.et al.The exid dataset: A real-world trajectory dataset of highly interactive highway scenarios in germany. In2022 IEEE Intelligent Vehicles Symposium (IV), 958–964 (IEEE, 2022)

    2022

  49. [49]

    In 2020 IEEE Intelligent Vehicles Symposium (IV), 1929–1934 (IEEE, 2020)

    Bock, J.et al.The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections. In 2020 IEEE Intelligent Vehicles Symposium (IV), 1929–1934 (IEEE, 2020)

    2020

  50. [50]

    In2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC), 2471–2478 (IEEE, 2022)

    Xu, Y .et al.Sind: A drone dataset at signalized intersection in china. In2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC), 2471–2478 (IEEE, 2022)

    2022

  51. [51]

    Li, G., Jiao, Y ., Calvert, S. C. & van Lint, J. H. Lateral conflict resolution data derived from argoverse-2: Analysing safety and efficiency impacts of autonomous vehicles at intersections.Transp. Res. Part C: Emerg. Technol.167, 104802 (2024)

    2024

  52. [52]

    & Risser, R

    Almqvist, S., Hyden, C. & Risser, R. Use of speed limiters in cars for increased safety and a better environment. Transp. Res. Rec.(1991). 54.Jiao, Y . A fast calculation of two-dimensional time-to-collision. GitHub repository (2023)

    1991

  53. [53]

    & Chen, Y .-H

    Huang, Y .-L. & Chen, Y .-H. Estimating intersections’ near-crash conflicts with the drone-based image-recording data.Traffic Saf. Res.9, e000084–e000084 (2025)

    2025

  54. [54]

    arXiv preprint arXiv:2510.27333(2025)

    Cheng, H.et al.Modified-emergency index (mei): A criticality metric for autonomous driving in lateral conflict. arXiv preprint arXiv:2510.27333(2025)

    work page arXiv 2025

  55. [55]

    L., Shin, B

    Allen, B. L., Shin, B. T. & Cooper, P. J. Analysis of traffic conflicts and collisions. Tech. Rep., Texas Transportation Institute (1978)

    1978

  56. [56]

    & Hydén, C

    Laureshyn, A., Svensson, Å. & Hydén, C. Evaluation of traffic safety, based on micro-level behavioural data: Theoretical framework and first implementation.Accid. Analysis & Prev.42, 1637–1646 (2010). 21/23

    2010

  57. [57]

    Cheng, H.et al.Emergency index (ei): A two-dimensional surrogate safety measure considering vehicles’ interaction depth.Transp. Res. Part C: Emerg. Technol.171, 104981 (2025)

    2025

  58. [58]

    C., van Cranenburgh, S

    Jiao, Y ., Calvert, S. C., van Cranenburgh, S. & van Lint, H. Learning collision risk proactively from naturalistic driving data at scale.Nat. Mach. Intell.1–14 (2026)

    2026

  59. [59]

    Polychronopoulos, A., Tsogas, M., Amditis, A. J. & Andreone, L. Sensor fusion for predicting vehicles’ path for collision avoidance systems.IEEE Transactions on Intell. Transp. Syst.8, 549–562 (2007)

    2007

  60. [60]

    & Wanielik, G

    Schubert, R., Richter, E. & Wanielik, G. Comparison and evaluation of advanced motion models for vehicle tracking. In2008 11th international conference on information fusion, 1–6 (IEEE, 2008)

    2008

  61. [61]

    In2011 IEEE intelligent vehicles symposium (IV), 534–539 (IEEE, 2011)

    Schubert, R.et al.Empirical evaluation of vehicular models for ego motion estimation. In2011 IEEE intelligent vehicles symposium (IV), 534–539 (IEEE, 2011)

    2011

  62. [62]

    & Lee, S

    Baek, M., Jeong, D., Choi, D. & Lee, S. Vehicle trajectory prediction and collision warning via fusion of multisensors and wireless vehicular communications.Sensors20, 288 (2020). 65.Brow, G. R. Traffic conflicts for road user safety studies.Can. J. Civ. Eng.21, 1–15 (1994)

    2020

  63. [63]

    & Sayed, T

    El-Basyouny, K. & Sayed, T. Safety performance functions using traffic conflicts.Saf. science51, 160–164 (2013)

    2013

  64. [64]

    & Imprialou, M

    Papadoulis, A., Quddus, M. & Imprialou, M. Evaluating the safety impact of connected and autonomous vehicles on motorways.Accid. Analysis & Prev.124, 12–22 (2019). 68.Sayed, T. & Zein, S. Traffic conflict standards for intersections.Transp. Plan. Technol.22, 309–323 (1999)

    2019

  65. [65]

    & Zhao, J

    Zhang, G., Chen, J. & Zhao, J. Safety performance evaluation of a three-leg unsignalized intersection using traffic conflict analysis.Math. problems engineering2017, 2948750 (2017)

    2017

  66. [66]

    & Xing, L

    Li, Y ., Li, Z., Wang, H., Wang, W. & Xing, L. Evaluating the safety impact of adaptive cruise control in traffic oscillations on freeways.Accid. Analysis & Prev.104, 137–145 (2017)

    2017

  67. [67]

    Rahman, M. S. & Abdel-Aty, M. Longitudinal safety evaluation of connected vehicles’ platooning on express- ways.Accid. Analysis & Prev.117, 381–391 (2018)

    2018

  68. [68]

    S., Abdel-Aty, M., Lee, J

    Rahman, M. S., Abdel-Aty, M., Lee, J. & Rahman, M. H. Safety benefits of arterials’ crash risk under connected and automated vehicles.Transp. Res. Part C: Emerg. Technol.100, 354–371 (2019)

    2019

  69. [69]

    & Liu, P

    Wang, C., Xie, Y ., Huang, H. & Liu, P. A review of surrogate safety measures and their applications in connected and automated vehicles safety modeling.Accid. Analysis & Prev.157, 106157 (2021). Acknowledgements This work was supported by the National Natural Science Foundation of China (Grant Nos. 52572497 and 52572477), the National Key R&D Program of ...

    2021