Predictive Risk Analysis and Safe Trajectory Planning for Intelligent and Connected Vehicles

Chaoyi Chen; Guangwei Wang; Jianqiang Wang; Keqiang Li; Mengchi Cai; Qingwen Meng; Qing Xu; Ying Liu; Zeyu Han

REVIEW 6 major objections 5 minor 24 references

Predictive Risk Analysis and Safe Trajectory Planning for Intelligent and Connected Vehicles

T0 review · 6 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A predictive risk field that looks ahead over future time slices can detect driving hazards more reliably than current-instant time-to-collision, allowing an intelligent connected vehicle to plan a safe, real-time trajectory.

desk verdict A plausible predictive risk-field planner that is undercut by an ill-defined field equation and an uncheckable accuracy claim. read the letter →

arxiv 2506.23999 v1 pith:GA67SMYN submitted 2025-06-30 cs.RO

Zeyu Han , Mengchi Cai , Chaoyi Chen , Qingwen Meng , Guangwei Wang , Ying Liu , Qing Xu , Jianqiang Wang

show 1 more author

Keqiang Li

This is my paper · ORCID

classification cs.RO

keywords trajectorypredictionriskfieldsafeplanningintelligentandconnectedvehiclesspatiotemporalanalysistime-to-collisionLSTMself-attentionautonomousdriving

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Most driving-risk measures look only at the present instant: how close a vehicle is, or how soon a collision would occur if nothing changed. This paper tries to establish that an autonomous vehicle should instead assess risk over future time slices, using predicted trajectories of surrounding objects. It builds a spatiotemporal risk field that combines the current scene with predicted future positions, then plans a trajectory that stays within safe risk levels. The authors argue this makes the vehicle less conservative while still safe, and report that their risk-field method detects hazards with 95.83% accuracy versus 62.50% for time-to-collision, at an average 26 ms per risk analysis.

What carries the argument

The central object is the spatiotemporal-discretised predictive risk field: a time-indexed grid of risk values built from predicted trajectories, evaluated every 0.5 s. Each dynamic object contributes potential energy $pE_i^d(x,y)$, shaped by a virtual mass $M_i^v=m\cdot T\cdot(\alpha v^\beta+\gamma)$ and distance, and kinetic energy $kE_i^d(x,y)$, shaped by relative velocity and a weighted Frenet distance $D_w(x,y)$ that makes the field extend ahead of the object in its direction of motion. The field's role is to convert prediction outputs into a risk map that the planner can check slice by slice, so that each path segment's desired speed is capped by the maximum risk encountered in that segment.

What would settle it

Run the risk-field model on a held-out dataset of crash and near-crash events with known outcomes and check whether its maximum risk over future time slices ranks collision-bound scenes above safe scenes more accurately than time-to-collision; if it does not, the claimed detection advantage is falsified. A second test: compare trajectories from this framework against a planner that uses only current-instant risk on the same scenarios; the predictive planner should avoid more collisions or reduce unnecessary braking if the central claim is correct.

Watch

Extended reading notes

Core claim

The central claim is that a spatiotemporal-discretised predictive risk field, computed from locally predicted future trajectories, is a more informative basis for safe trajectory planning than current-instant risk measures such as time-to-collision. In this framework each object's risk is represented by potential field energy $pE_i^d(x,y)$ and kinetic field energy $kE_i^d(x,y)$, with future fields evaluated at 0.5 s intervals; static objects such as lane lines add potential energy only. Candidate paths are sampled, sliced into future time segments, scored by the maximum risk in the corresponding future field, and the path with the highest feasible desired speed is selected. The paper reports that this predictive risk assessment achieves 95.83% accuracy against 62.50% for TTC, runs in about 26 ms, and supports real-vehicle trajectory planning that avoids unnecessary braking.

Load-bearing premise

The load-bearing premise is that the hand-crafted risk-field equations, with roughly nine parameters and no reported calibration, genuinely measure collision risk; if high field values do not align with real danger, the safety guarantee for the planned trajectories does not follow.

Editorial extensions

If this is right

A vehicle using this framework can perceive a developing hazard in a future risk field and begin slowing before a time-to-collision threshold would trigger braking.
The three modules run in real time on vehicle hardware when the GPU-based parallel risk-field computation is used: 65.96 ms for prediction, 26.00 ms for risk analysis, and 20.05 ms for planning.
In the reported simulations, the planner handles left turns, straight driving, right turns, U-turns, and overtaking while keeping traffic efficiency at least 90% of a human driver.
The 95.83% hazard-detection accuracy, versus 62.50% for TTC, implies fewer missed risks and fewer false braking events in the tested scenarios.

Reading between the lines

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

If the risk-field equations remain valid beyond the tested road types and sensor conditions, the same time-sliced risk volume could be shared across a fleet of connected vehicles, since each vehicle's predicted trajectory could update a common grid.
The comparison against TTC is a single baseline; a natural extension would be to test against other current-instant surrogates such as time headway or post-encroachment time, and against predictive planners, to isolate how much of the gain comes from prediction versus the field's shape.
The 26 ms per risk analysis suggests the grid computation has headroom to scale to denser scenes or finer time slices, which a follow-up study could verify by measuring runtime as the object count grows.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: the architecture—prediction feeding a spatiotemporal risk field, then using that field to plan speeds—is coherent engineering, but the validation does not back the headline numbers, and Eq. (4) has a mathematical problem that needs fixing.

What is actually new: combining a risk-aware LSTM (prior work from the same group) with a risk-field planner, discretized in time. The ablation in Table 1 shows the risk module and self-attention both reduce ADE/FDE on their collected dataset. The path-slicing planner is a reasonable way to keep the spatiotemporal risk computation tractable, and the CUDA implementation reports 26 ms per risk analysis, which is plausible for real-time use.

The soft spots are in the validation. Eq. (4) defines kinetic field energy as v_r^2/[2 M_v D_w (1 - D_w/(v_r t))]. For any point with D_w > v_r t the denominator is negative, so the kinetic energy is negative; as D_w approaches v_r t it diverges. The paper says smoothing is applied but never writes the smoothed formula. As printed, the risk matrix is not generally finite or positive, and Algorithm 1 selects speeds based on the maximum risk in each slice, so an undefined field can produce arbitrary planning decisions.

Second, the central experiment in §4.1.2 reports 95.83% accuracy for the risk field versus 62.50% for TTC, but 'accuracy' is never defined. There is no list of scenarios, no trial count, no ground-truth labels, and no risk threshold. That number is the paper's main evidence that the risk field reflects true hazard, but it is not independently checkable.

Third, the trajectory prediction results have no error bars or significance tests, and the vehicle experiments are qualitative. The risk field parameters (nine in Eqs. (2)-(4) plus several more) are not calibrated and get no sensitivity analysis. There is also some circularity in using the risk field to improve prediction and then using prediction to build the field, but that is not automatically fatal; the missing accuracy metric is the bigger problem.

Who this is for: researchers working on risk-aware motion planning who want to see one way to stitch prediction into a risk field. The idea is worth a referee's time, but the paper needs substantial revision: fix Eq. (4), define and report the risk accuracy metric, add error bars and external baselines for prediction, and ideally release code/data. I would not accept it as is.

Referee Report

6 major / 5 minor

Summary. The paper proposes a three-module framework for intelligent and connected vehicles: a local risk-aware LSTM trajectory predictor with self-attention, a spatiotemporal-discretized predictive risk field that combines potential and kinetic energy terms, and a sampling-based safe trajectory planner that selects speeds from risk-field maxima along path slices. The authors report simulation results for trajectory prediction (ADE/FDE), a risk-analysis experiment claiming 95.83% accuracy versus 62.50% for TTC, qualitative planning simulations, and a vehicle experiment with module-level runtime measurements. The central claim is that the framework is effective and real-time, and that incorporating predicted trajectories into risk assessment improves safety over current-information-only methods.

Significance. If substantiated, the proposed framework would be a useful engineering contribution: it connects trajectory prediction, predictive risk assessment, and trajectory planning in a single modular pipeline, and the reported runtimes (26 ms for risk analysis, 65.96 ms for prediction, 20.05 ms for planning) suggest real-time feasibility. The authors are also to be credited for including ablation-style comparisons in the prediction module and for attempting vehicle experiments. However, the significance cannot currently be assessed because the risk field itself is not a well-defined, calibrated, or externally validated quantity, and the headline accuracy numbers and ADE/FDE improvements lack statistical grounding. The paper therefore does not yet support its claimed confirmation of efficacy and safety.

major comments (6)

[Sec. 3.2, Eq. (4)] The kinetic field energy formula in Eq. (4) contains the factor (1 - D_w/(v_r t)) in the denominator. For any point with D_w > v_r t, this factor is negative, producing negative kinetic energy; as D_w approaches v_r t from below, the energy diverges. The text states that 'potential energy and logarithmic smoothing' are applied to restrict and smooth the field, but the actual smoothed formula is never given. Since Algorithm 1 selects desired speeds by taking the maximum risk in each slice, undefined or negative risk values can lead to arbitrary planning decisions. The authors must provide the complete, well-defined expression for the kinetic field and show that it is finite and nonnegative (or explain how negative/divergent values are handled in the risk matrices).
[Sec. 4.1.2] The reported risk-analysis accuracy of 95.83% for the risk field and 62.50% for TTC is unsupported because the term 'accuracy' is never defined. The paper does not specify the ground-truth hazard labels, the risk threshold used to convert field values into binary risk predictions, the number and type of hazardous scenarios, the number of trials, or the confusion matrix. Without this information, the claimed 53.33% relative improvement over TTC cannot be independently checked, and the result does not establish that the risk field reflects true hazard. A proper evaluation protocol with an external ground truth should be provided.
[Sec. 4.1.1, Table 1] The trajectory prediction results in Table 1 are reported as point values of ADE/FDE with no error bars, no number of test trajectories, and no significance tests. The differences between the original model and the ablation without risk assessment are small (e.g., 0.49 vs. 0.53 at 20 frames, 0.82 vs. 0.85 at 40 frames), and could be within run-to-run or dataset variability. The claim that the risk assessment and self-attention modules 'observably improve' performance requires variance estimates and statistical testing.
[Sec. 3.2, Eqs. (2)-(4)] The risk field model introduces at least nine parameters (k, ra, k1, Emax, T, alpha, beta, gamma, t) plus ks and kappa, with no derivation, calibration procedure, or sensitivity analysis. The risk field is the load-bearing component of the entire safety argument: planning decisions are based on its values. Without calibration to real data and an analysis of how sensitive the outputs are to parameter choices, the claimed safety properties of the planner are unsupported.
[Sec. 4.2] The vehicle experiments are qualitative: only one scenario (a large truck in an adjacent lane) is described, with a figure showing a speed reduction, and the text reports no quantitative safety metrics such as minimum clearance, TTC distribution, or constraint violations. The conclusion that the algorithm is 'safe and comfortable' is stronger than the evidence supports. The authors should report quantitative results across multiple scenarios, ideally with comparisons to a baseline planner.
[Sec. 3.1 and Sec. 3.2] There is a potential circularity in the validation chain: the risk field is used as input to the LSTM trajectory prediction, and the predicted trajectories are then used to construct the predictive risk field. The risk-analysis accuracy experiment in Sec. 4.1.2 must therefore specify a ground truth that is independent of the risk field itself. As written, the paper does not explain how the ground-truth labels for the 95.83% accuracy figure were obtained, leaving open the possibility that the same risk field generated both the predictions and the labels.

minor comments (5)

[Throughout] There are several typos and formatting issues: 'detailedly' should be 'in detail', 'future moments..' has a double period, 'T able' should be 'Table', and 'the the maximum risk' appears in Sec. 3.3. These should be corrected.
[Sec. 3.2] The sentence 'Most existing driving risk analysis algorithms[] calculate...' contains an empty bracket where a citation or reference should be.
[Sec. 3.3, Algorithm 1] The notation {P athi}n i=1 is visually awkward and the variable S in the PathSlicing call is not defined in the input/initialization section; please clarify what S denotes (it appears to be the ego vehicle's current speed).
[Sec. 4.1.2] The claim that the risk field result is 'approximately 53.33% better relative to the traditional TTC' is arithmetically confusing: (95.83 - 62.50)/62.50 = 53.33%, but this relative improvement is not a meaningful comparison without the underlying sample sizes. Please report absolute counts and confidence intervals.
[Sec. 4.2, Table 2] The authors state that the three modules can operate in parallel, but no evidence of parallel execution on the test hardware is provided; the summed module times (112 ms) exceed the reported per-module times, so clarify whether the 26 ms and 20.05 ms figures are measured in isolation or in the integrated pipeline.

Circularity Check

0 steps flagged · score 0.0 of 10

No constructional circularity: the risk field, trajectory predictor, and planner are distinct modules; the main gaps are an undefined accuracy metric and an ill-conditioned risk formula, which are correctness issues rather than circular derivation.

full rationale

The claimed derivation chain is not circular by construction. Section 3.1 builds an LSTM trajectory predictor whose input is history positions plus surrounding-risk features: 'Xi indicates the ith frame of history information, including 2-dimension positions and 9-dimension surrounding risks. The risk of each surrounding position is calculated by other objects refer to [23, 24].' Section 3.2 then forms current and future risk fields from predicted positions and velocities: 'The future risk fields are calculated in a similar manner based on the trajectory prediction results.' Section 3.3 plans by evaluating sampled paths against the risk matrices (Algorithm 1). These are nested modules, not a single equation whose output is its own input; the LSTM ablation against a 'without risk assessment' model provides an empirical control for the risk feature, and using the same field in prediction and planning is a design consistency rather than a theorem that reduces to its assumptions. The only same-group citations ([23,24] supply the risk-field feature) are not load-bearing alternatives-forbidding evidence; the paper reports its own ablations and simulations. The serious weaknesses are not circularity: Section 4.1.2 reports that 'The accuracy of the risk field method is 95.83%, whereas the traditional TTC achieves an accuracy of 62.50%' without defining the ground-truth hazard labels, scenario counts, or risk thresholds, and Eq. (4) as printed can become negative or diverge before the promised smoothing ('We also apply the potential energy and logarithmic smoothing to restrict and smooth the kinetic field energy'). These are evidentiary and mathematical correctness gaps that need repair, but they do not make any prediction equivalent to its inputs by construction.

Assumptions & free parameters 12 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a risk field model with twelve free parameters and three key assumptions. No new physical entities are introduced. The additivity and functional forms are adopted from prior work but extended here without calibration, so the burden on these assumptions is high.

free parameters (12)

k
Scaling factor in potential field energy, Eq. (2).
ra
Parameter in potential field energy, Eq. (2).
k1
Exponent in distance term of potential field energy, Eq. (2).
Emax
Maximum energy normalization in potential field energy, Eq. (2).
T
Time factor in virtual mass definition, Eq. (2).
alpha
Coefficient in virtual mass velocity term, Eq. (2).
beta
Exponent on velocity in virtual mass, Eq. (2).
gamma
Constant term in virtual mass, Eq. (2).
w
Width of object in weighted distance, Eq. (3).
t
Time threshold in kinetic field energy, Eq. (4).
ks
Scaling factor for static object field, Eq. (5).
kappa
Ratio parameter for static object width, Eq. (5).

assumptions (3)

domain assumption Risk fields of multiple objects are additive (Eq. 1).
The total risk is the sum of individual object fields; no evidence is given that this linear superposition holds.
ad hoc to paper The functional forms of potential and kinetic field energies (Eqs. 2 and 4) are correct.
No derivation or empirical validation is provided for these hand-crafted formulas.
domain assumption LSTM-predicted future trajectories are accurate enough to compute future risk fields.
The risk analysis depends on prediction quality, which is only weakly validated with point estimates and no error bars.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Predictive Risk Analysis and Safe Trajectory Planning for Intelligent and Connected Vehicles." pith.science (2026). https://pith.science/paper/GA67SMYN

@misc{pith2026250623999,
  author       = {Pith},
  title        = {Pith review of: Predictive Risk Analysis and Safe Trajectory Planning for Intelligent and Connected Vehicles},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GA67SMYN}},
  note         = {Machine review of arXiv:2506.23999}
}

read the original abstract

The safe trajectory planning of intelligent and connected vehicles is a key component in autonomous driving technology. Modeling the environment risk information by field is a promising and effective approach for safe trajectory planning. However, existing risk assessment theories only analyze the risk by current information, ignoring future prediction. This paper proposes a predictive risk analysis and safe trajectory planning framework for intelligent and connected vehicles. This framework first predicts future trajectories of objects by a local risk-aware algorithm, following with a spatiotemporal-discretised predictive risk analysis using the prediction results. Then the safe trajectory is generated based on the predictive risk analysis. Finally, simulation and vehicle experiments confirm the efficacy and real-time practicability of our approach.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 23 canonical work pages

[1]

Engineering 7(9), 1313–1325 (2021)

Wang, J., Huang, H., Li, K., Li, J.: Towards the unified principles for level 5 autonomous vehicles. Engineering 7(9), 1313–1325 (2021)

work page 2021
[2]

IEEE Transactions on Intelligent Transportation Systems (2023)

Wang, Y., Jiang, J., Li, S., Li, R., Xu, S., Wang, J., Li, K.: Decision-making driven by driver intelligence and environment reasoning for high-level autonomous vehicles: a survey. IEEE Transactions on Intelligent Transportation Systems (2023)

work page 2023
[3]

Automotive Innova- tion, 1–31 (2025)

Yang, Y., Negash, N.M., Yang, J.: Recent advances in interactive driving of autonomous vehicles: Comprehensive review of approaches. Automotive Innova- tion, 1–31 (2025)

work page 2025
[4]

IEEE Intelligent Transportation Systems Magazine 13(4), 71–91 (2020)

Li, Y., Li, K., Zheng, Y., Morys, B., Pan, S., Wang, J.: Threat assess- ment techniques in intelligent vehicles: A comparative survey. IEEE Intelligent Transportation Systems Magazine 13(4), 71–91 (2020)

work page 2020
[5]

Transportation Science 10(2), 125–148 (1976)

Branston, D.: Models of single lane time headway distributions. Transportation Science 10(2), 125–148 (1976)

work page 1976
[6]

IEEE Transactions on Intelligent Transportation Systems 23(11), 21190–21200 (2022)

Li, Y., Lv, Q., Zhu, H., Li, H., Li, H., Hu, S., Yu, S., Wang, Y.: Variable time headway policy based platoon control for heterogeneous connected vehicles with external disturbances. IEEE Transactions on Intelligent Transportation Systems 23(11), 21190–21200 (2022)

work page 2022
[7]

Perception 5(4), 437–459 (1976)

Lee, D.N.: A theory of visual control of braking based on information about time-to-collision. Perception 5(4), 437–459 (1976)

work page 1976
[8]

Transportation letters 12(7), 491–500 (2020)

Nadimi, N., Ragland, D.R., Mohammadian Amiri, A.: An evaluation of time-to- collision as a surrogate safety measure and a proposal of a new method for its application in safety analysis. Transportation letters 12(7), 491–500 (2020)

work page 2020

Show all 24 references

[9]

In: 2009 IEEE Intelligent Vehicles Symposium, pp

Chang, T.-H., Chou, C.-J.: Rear-end collision warning system on account of a rear-end monitoring camera. In: 2009 IEEE Intelligent Vehicles Symposium, pp. 913–917 (2009). IEEE 12

2009
[10]

Automotive Innovation 6(3), 466–480 (2023)

Wang, Y., Cao, X., Hu, Y.: A trajectory planning method of automatic lane change based on dynamic safety domain. Automotive Innovation 6(3), 466–480 (2023)

2023
[11]

Journal of Intelligent and Connected Vehicles (2024)

Shi, J., Luo, Y., Li, P., Wang, J., Li, K.: Collaborative multi-lane on-ramp merg- ing strategy for connected and automated vehicles using dynamic conflict graph. Journal of Intelligent and Connected Vehicles (2024)

2024
[12]

IEEE Transactions on intelligent transportation systems 12(4), 942–953 (2011)

Greene, D., Liu, J., Reich, J., Hirokawa, Y., Shinagawa, A., Ito, H., Mikami, T.: An efficient computational architecture for a collision early-warning sys- tem for vehicles, pedestrians, and bicyclists. IEEE Transactions on intelligent transportation systems 12(4), 942–953 (2011)

2011
[13]

Journal of Intelligent and Connected Vehicles7(3), 205–218 (2024)

Chen, J., Zhang, W., Liu, M., Wang, X., Gong, J., Li, J., Li, B., Xu, J.: Spectrum quantification-based safety efficiency evaluation of autonomous vehicle under ran- dom cut-in scenarios. Journal of Intelligent and Connected Vehicles7(3), 205–218 (2024)

2024
[14]

Automotive Innovation, 1–12 (2025)

Liu, R., Zhao, X., Zhu, X., Ma, J.: Comparison between cut-in performance model of ece r157 and behaviors of chinese drivers. Automotive Innovation, 1–12 (2025)

2025
[15]

In: Proc

Krogh, B.: A generalized potential field approach to obstacle avoidance control. In: Proc. SME Conf. on Robotics Research: The Next Five Years and Beyond, Bethlehem, PA, 1984, pp. 11–22 (1984)

1984
[16]

Rossetter, E.J., Gerdes, J.C.: Lyapunov based performance guarantees for the potential field lane-keeping assistance system (2006)

2006
[17]

In: Icctp 2011: Towards Sustainable Transportation Systems, pp

Ni, D.: A unified perspective on traffic flow theory, part i: the field theory. In: Icctp 2011: Towards Sustainable Transportation Systems, pp. 4227–4243 (2011)

2011
[18]

IF AC Proceedings Volumes 46(21), 328–334 (2013)

Matsumi, R., Raksincharoensak, P., Nagai, M.: Autonomous braking control sys- tem for pedestrian collision avoidance by using potential field. IF AC Proceedings Volumes 46(21), 328–334 (2013)

2013
[19]

IEEE Transactions on Intelligent Transportation Systems 16(4), 2203–2214 (2015)

Wang, J., Wu, J., Li, Y.: The driving safety field based on driver–vehicle–road interactions. IEEE Transactions on Intelligent Transportation Systems 16(4), 2203–2214 (2015)

2015
[20]

In: 2021 IEEE International Conference on Robotics and Biomimetics (ROBIO), pp

Liu, J., Mao, X., Fang, Y., Zhu, D., Meng, M.Q.-H.: A survey on deep-learning approaches for vehicle trajectory prediction in autonomous driving. In: 2021 IEEE International Conference on Robotics and Biomimetics (ROBIO), pp. 978–985 (2021). IEEE

2021
[21]

In: 2018 IEEE Intelligent Vehicles Symposium (IV), pp

Park, S.H., Kim, B., Kang, C.M., Chung, C.C., Choi, J.W.: Sequence-to-sequence prediction of vehicle trajectory via lstm encoder-decoder architecture. In: 2018 IEEE Intelligent Vehicles Symposium (IV), pp. 1672–1678 (2018). IEEE 13

2018
[22]

arXiv preprint arXiv:1803.02155 (2018)

Shaw, P., Uszkoreit, J., Vaswani, A.: Self-attention with relative position repre- sentations. arXiv preprint arXiv:1803.02155 (2018)

2018 arXiv
[23]

Transportation research part C: emerging technologies 72, 306–324 (2016)

Wang, J., Wu, J., Zheng, X., Ni, D., Li, K.: Driving safety field theory modeling and its application in pre-collision warning system. Transportation research part C: emerging technologies 72, 306–324 (2016)

2016
[24]

IEEE Transactions on Vehicular Technology (2023) 14

Wang, Y., Wang, J., Jiang, J., Xu, S., Wang, J.: Sa-lstm: A trajectory prediction model for complex off-road multi-agent systems considering situation awareness based on risk field. IEEE Transactions on Vehicular Technology (2023) 14

2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Engineering 7(9), 1313–1325 (2021)

Wang, J., Huang, H., Li, K., Li, J.: Towards the unified principles for level 5 autonomous vehicles. Engineering 7(9), 1313–1325 (2021)

work page 2021

[2] [2]

IEEE Transactions on Intelligent Transportation Systems (2023)

Wang, Y., Jiang, J., Li, S., Li, R., Xu, S., Wang, J., Li, K.: Decision-making driven by driver intelligence and environment reasoning for high-level autonomous vehicles: a survey. IEEE Transactions on Intelligent Transportation Systems (2023)

work page 2023

[3] [3]

Automotive Innova- tion, 1–31 (2025)

Yang, Y., Negash, N.M., Yang, J.: Recent advances in interactive driving of autonomous vehicles: Comprehensive review of approaches. Automotive Innova- tion, 1–31 (2025)

work page 2025

[4] [4]

IEEE Intelligent Transportation Systems Magazine 13(4), 71–91 (2020)

Li, Y., Li, K., Zheng, Y., Morys, B., Pan, S., Wang, J.: Threat assess- ment techniques in intelligent vehicles: A comparative survey. IEEE Intelligent Transportation Systems Magazine 13(4), 71–91 (2020)

work page 2020

[5] [5]

Transportation Science 10(2), 125–148 (1976)

Branston, D.: Models of single lane time headway distributions. Transportation Science 10(2), 125–148 (1976)

work page 1976

[6] [6]

IEEE Transactions on Intelligent Transportation Systems 23(11), 21190–21200 (2022)

Li, Y., Lv, Q., Zhu, H., Li, H., Li, H., Hu, S., Yu, S., Wang, Y.: Variable time headway policy based platoon control for heterogeneous connected vehicles with external disturbances. IEEE Transactions on Intelligent Transportation Systems 23(11), 21190–21200 (2022)

work page 2022

[7] [7]

Perception 5(4), 437–459 (1976)

Lee, D.N.: A theory of visual control of braking based on information about time-to-collision. Perception 5(4), 437–459 (1976)

work page 1976

[8] [8]

Transportation letters 12(7), 491–500 (2020)

Nadimi, N., Ragland, D.R., Mohammadian Amiri, A.: An evaluation of time-to- collision as a surrogate safety measure and a proposal of a new method for its application in safety analysis. Transportation letters 12(7), 491–500 (2020)

work page 2020