arxiv: 2605.05050 · v1 · submitted 2026-05-06 · 📡 eess.SY · cs.LG· cs.SY

Recognition: unknown

Kinematic Discriminants of Deceleration Behavior Modes in Car-Following: Evidence from NGSIM Trajectory Data

Eni Solomon Laughter

Authors on Pith no claims yet

Pith reviewed 2026-05-08 16:14 UTC · model grok-4.3

classification 📡 eess.SY cs.LGcs.SY

keywords car-followingdeceleration behaviorkinematic featuresvisual loominggap-closing rateNGSIM dataK-means clusteringdriver cue prioritization

0 comments

The pith

Gap-closing rate and visual looming swap dominance depending on braking intensity in car-following.

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

The paper analyzes over a million car-following segments from the NGSIM dataset to separate what kinematic information is available in the scene from what drivers actually use to distinguish their own deceleration patterns. K-means clustering on six features, followed by ANOVA effect-size ranking, shows that hard braking events separate best by gap-closing rate while moderate braking events separate best by visual looming; spacing headway contributes almost nothing in either case. The stricter deceleration threshold produces three distinct modes whereas the looser threshold collapses them to two, demonstrating that the choice of braking definition itself alters the inferred behavior. These results directly contradict the long-standing emphasis on spacing in classic car-following models and supply concrete candidates for which cues should be prioritized in driver models.

Core claim

In 1,060,119 valid NGSIM car-following observations, K-means clustering on kinematic features identifies behavioral modes whose separation is dominated by gap-closing rate (eta-squared 0.715) under a strict -0.5 m/s² threshold and by visual looming (eta-squared 0.574) under a permissive -0.3 m/s² threshold, while spacing headway remains negligible (eta-squared ≤ 0.014) in both regimes.

What carries the argument

Two-stage framework that first extracts six kinematic variables from trajectories, detects deceleration events at chosen thresholds, applies K-means to reveal modes, and then uses one-way ANOVA with eta-squared to rank each variable's ability to discriminate those modes.

If this is right

Threshold choice for defining a braking event determines whether two or three distinct deceleration modes are recovered.
Hard and moderate braking recruit different primary perceptual cues, so models assuming a single fixed cue across intensities are incomplete.
Spacing headway contributes negligible discriminative information once gap-closing rate and looming are included.
ADAS warning algorithms and autonomous-vehicle longitudinal controllers can be made more accurate by weighting cues according to expected deceleration intensity.
Traditional spacing-centered car-following models require revision to accommodate intensity-dependent cue prioritization.

Where Pith is reading between the lines

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

Warning systems could adapt their urgency thresholds dynamically by monitoring whether the current closing rate or looming rate is the stronger predictor at the observed intensity.
The same two-stage availability-versus-utilization test could be applied to acceleration or lane-change events to check whether cue dominance also shifts with maneuver intensity.
If simulator studies confirm that human reaction times track the same eta-squared ordering, the kinematic ranking supplies a direct mapping from measurable scene variables to driver response.
The negligible role of spacing suggests that purely distance-based safety margins in traffic flow models may systematically mispredict braking timing.

Load-bearing premise

The K-means clusters on the chosen kinematic features correspond to genuine driver behavioral modes rather than artifacts of feature selection or the specific deceleration thresholds used.

What would settle it

Re-running the identical pipeline on the same NGSIM segments but with a different set of kinematic features or with thresholds shifted by 0.1 m/s² yields either the same dominance ordering or a spacing headway eta-squared above 0.1.

read the original abstract

Gap-closing rate and visual looming swap discriminative dominance depending on deceleration intensity - a finding that reconciles a long-standing conflict in the car-following literature and challenges spacing-centered assumptions in traditional driver behavior models. This study presents a two-stage analytical framework that distinguishes between information availability (kinematic variables measurable in the environment) and information utilization (variables that demonstrably separate driver behavioral patterns), applied to 1,060,119 valid car-following observations from the NGSIM trajectory dataset (2,932 vehicles). Six kinematic features are extracted, and deceleration events are detected under two threshold conditions (-0.5 m/s^2 and -0.3 m/s^2). K-means clustering identifies behavioral modes, and one-way ANOVA with eta-squared effect sizes ranks each feature's discriminative power. Three key findings emerge: (1) threshold selection fundamentally shapes behavioral inference - the stricter threshold yields three interpretable modes while the permissive threshold collapses these to two; (2) hard braking prioritizes gap-closing rate (eta^2 = 0.715) while moderate braking emphasizes visual looming (eta^2 = 0.574); and (3) spacing headway is negligible (eta^2 <= 0.014) across both thresholds. These findings provide empirically grounded candidates for perceptual cue prioritization and have direct implications for ADAS warning system design and autonomous vehicle control.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note private letter to a colleague

This NGSIM analysis shows gap-closing rate dominating hard-braking mode separation while visual looming leads for moderate braking, with spacing near zero, but the K-means plus ANOVA pipeline has validation gaps.

read the letter

The main thing here is that the paper reports a clear intensity-dependent swap in which kinematic feature best separates driver modes from over a million NGSIM observations: gap-closing rate leads for hard braking (eta-squared 0.715) and visual looming for moderate (0.574), while spacing headway stays negligible across both. That specific ranking and the threshold sensitivity (three modes vs two) is the new empirical piece that could help settle older conflicts in the car-following literature. The work does a straightforward job of extracting six standard kinematic features from public trajectory data, applying K-means to label modes, and then using one-way ANOVA with effect sizes to rank discriminative power. Reporting concrete numbers and noting how threshold choice alters the mode count gives readers something testable rather than vague claims. The soft spots sit in the methods. Running K-means on the same six features that later get ranked by ANOVA means some dominance is built in, and the absence of silhouette scores, alternative clustering checks, or external labels leaves open whether the modes reflect real information-use patterns or just algorithmic splits. Repeated observations per vehicle also break the independence assumption in ANOVA, which can inflate the eta-squared values. The chosen deceleration cutoffs are reasonable but arbitrary, and sensitivity to them is shown but not fully explored. This is useful for transportation engineers and modelers who need prioritized cues for ADAS or mixed-traffic simulations. A reader working on perceptual driver models would get concrete guidance on what to emphasize at different intensities. I would send it to peer review. The public data and quantitative claims are worth a referee's time to check the clustering stability and repeated-measures handling.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a two-stage framework applied to 1,060,119 NGSIM car-following observations from 2,932 vehicles. Six kinematic features are extracted; deceleration events are identified under two thresholds (-0.5 m/s² and -0.3 m/s²); K-means clustering identifies behavioral modes; and one-way ANOVA with eta-squared ranks each feature's ability to separate the modes. The central claims are that threshold choice alters the number of modes (three vs. two), that gap-closing rate dominates discrimination for hard braking (eta²=0.715) while visual looming does for moderate braking (eta²=0.574), and that spacing headway is negligible (eta²≤0.014) in both cases. These results are interpreted as evidence of shifting information utilization that reconciles prior literature conflicts and challenges spacing-centered car-following models.

Significance. If the K-means-derived modes can be shown to reflect distinct driver information-utilization patterns rather than algorithmic partitions, the reported swap in feature dominance would provide a data-driven reconciliation of conflicting findings on gap-closing rate versus visual looming in the car-following literature and would support revised perceptual-cue assumptions in driver models. The large public dataset and explicit comparison of thresholds are strengths; however, the absence of cluster validation or repeated-measures correction limits the immediate implications for ADAS or autonomous-vehicle control design.

major comments (3)

[Methods (two-stage analytical framework and K-means clustering)] The same six kinematic features are used both to define the K-means clusters (behavioral modes) and to compute their discriminative power via one-way ANOVA eta-squared. This renders the high reported values (e.g., eta²=0.715 for gap-closing rate under the stricter threshold) expected by construction for whichever features best separate the partitions, weakening the claim that they demonstrate distinct information-utilization patterns.
[Statistical analysis (ANOVA)] The dataset contains 1,060,119 observations from only 2,932 vehicles, creating a repeated-measures structure. One-way ANOVA assumes independent observations; without a mixed-effects model or vehicle-level clustering, the eta-squared effect sizes and the claimed dominance swap are likely inflated.
[Deceleration event detection and threshold selection] Deceleration thresholds of -0.5 m/s² and -0.3 m/s² are used to separate hard and moderate braking, yet no a-priori justification or robustness checks (beyond the resulting mode counts) are provided. The reduction from three to two modes under the more permissive threshold indicates that the reported feature-dominance swap may be sensitive to these arbitrary cutoffs.

minor comments (2)

[Results (K-means clustering)] The manuscript should report cluster-validation metrics (silhouette scores, within-cluster sum of squares, or stability under bootstrap resampling) to support the interpretability of the K-means modes.
[Methods (feature extraction)] Details on feature extraction (exact formulas for visual looming and gap-closing rate) and any preprocessing or outlier removal steps are needed to allow replication.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment point by point below, providing clarifications and indicating revisions where the concerns are valid and can be addressed through additional analysis or discussion.

read point-by-point responses

Referee: The same six kinematic features are used both to define the K-means clusters (behavioral modes) and to compute their discriminative power via one-way ANOVA eta-squared. This renders the high reported values (e.g., eta²=0.715 for gap-closing rate under the stricter threshold) expected by construction for whichever features best separate the partitions, weakening the claim that they demonstrate distinct information-utilization patterns.

Authors: We acknowledge that employing the identical feature set for both unsupervised clustering and subsequent one-way ANOVA renders the eta-squared values descriptive of the resulting partitions by design. This is a standard post-hoc interpretation step in cluster analysis rather than an independent test of causal information utilization. The primary interpretive value lies in the relative dominance shift between gap-closing rate and visual looming when the same procedure is applied across different deceleration thresholds, which reveals changes in data structure. To strengthen the presentation, we will revise the methods and discussion sections to explicitly note this aspect of the approach and add supplementary analyses such as feature-subset clustering or internal validation metrics to better support the information-utilization interpretation. revision: partial
Referee: The dataset contains 1,060,119 observations from only 2,932 vehicles, creating a repeated-measures structure. One-way ANOVA assumes independent observations; without a mixed-effects model or vehicle-level clustering, the eta-squared effect sizes and the claimed dominance swap are likely inflated.

Authors: The referee correctly identifies the repeated-measures dependency arising from multiple observations per vehicle. Standard one-way ANOVA does not account for this intra-vehicle correlation, which can indeed inflate effect sizes. We agree this is a methodological limitation. In the revised manuscript we will re-estimate the discriminative power using linear mixed-effects models that include vehicle as a random effect, thereby providing adjusted effect sizes that properly reflect the hierarchical data structure. revision: yes
Referee: Deceleration thresholds of -0.5 m/s² and -0.3 m/s² are used to separate hard and moderate braking, yet no a-priori justification or robustness checks (beyond the resulting mode counts) are provided. The reduction from three to two modes under the more permissive threshold indicates that the reported feature-dominance swap may be sensitive to these arbitrary cutoffs.

Authors: The chosen thresholds align with values frequently employed in the car-following literature to demarcate moderate versus hard braking, yet the original submission did not cite these precedents or conduct broader sensitivity tests. The two-threshold comparison was presented as an initial demonstration of threshold sensitivity. We will revise the manuscript to include explicit literature references supporting the threshold selections and expand the robustness analyses with additional checks, such as incremental threshold variations and stability assessment of the observed dominance swap. revision: yes

Circularity Check

0 steps flagged

No significant circularity; analysis is descriptive clustering followed by post-hoc feature ranking on public data

full rationale

The paper applies standard K-means to six kinematic features extracted from NGSIM trajectories under two deceleration thresholds, then uses one-way ANOVA with eta-squared to rank the same features' ability to discriminate the resulting clusters. This is a conventional post-hoc interpretation step rather than a derivation that reduces to its inputs by construction. No equations, fitted parameters renamed as predictions, self-citations as load-bearing premises, or ansatzes smuggled via prior work are present. The reported dominance swap (gap-closing rate eta^2=0.715 vs. looming eta^2=0.574) and negligible spacing (eta^2<=0.014) are direct outputs of the observed data partitions, not tautological re-statements of the clustering inputs. The method is self-contained against external benchmarks and does not invoke uniqueness theorems or renamings of known results.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The paper relies on chosen thresholds as free parameters and standard assumptions about clustering validity and statistical tests, but introduces no new physical entities or ad-hoc inventions beyond the analytical framework.

free parameters (1)

deceleration thresholds = -0.5 m/s² and -0.3 m/s²
Selected to define hard and moderate braking events, affecting the number of modes identified.

axioms (2)

domain assumption K-means clustering can identify distinct driver behavioral modes from kinematic features
Invoked when applying clustering to separate deceleration events into interpretable groups.
standard math One-way ANOVA with eta-squared accurately ranks feature discriminative power
Standard statistical assumption that the test measures effect size correctly for the data.

pith-pipeline@v0.9.0 · 5556 in / 1619 out tokens · 59488 ms · 2026-05-08T16:14:54.733286+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 19 canonical work pages

[1]

Ali, E., Ahmed, M., & Yang, G. (2021). Normal and risky driving patterns identification in clear and rainy weather on freeway segments using vehicle kinematics trajectories and time series cluster analysis. IATSS RESEARCH, 45(1), 137–152. (WOS:000646969300016). https://doi.org/10.1016/j.iatssr.2020.07.002

work page doi:10.1016/j.iatssr.2020.07.002 2021
[2]

Cao, X., Wang, J., & Chen, C. (2020). A Modified Car-following Model Considering Traffic Density and Acceleration of Leading Vehicle. APPLIED SCIENCES-BASEL, 10(4). (WOS:000525287900075). https://doi.org/10.3390/app10041268

work page doi:10.3390/app10041268 2020
[3]

Chen, X., Zhu, M., Chen, K., Wang, P., Lu, H., Zhong, H., Han, X., Wang, X., & Wang, Y. (2023). FollowNet: A Comprehensive Benchmark for Car-Following Behavior Modeling. SCIENTIFIC DATA, 10(1). (WOS:001109374600006). https://doi.org/10.1038/s41597-023-02718-7

work page doi:10.1038/s41597-023-02718-7 2023
[4]

Da Lio, M., Mazzalai, A., Gurney, K., & Saroldi, A. (2018). Biologically Guided Driver Modeling: The Stop Behavior of Human Car Drivers. IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, 19(8), 2454–2469. (WOS:000440852900007). https://doi.org/10.1109/TITS.2017.2751526

work page doi:10.1109/tits.2017.2751526 2018
[5]

Durrani, U., Lee, C., & Shah, D. (2021). Predicting driver reaction time and deceleration: Comparison of perception-reaction thresholds and evidence accumulation framework. ACCIDENT ANALYSIS AND PREVENTION, 149. (WOS:000598055300002). https://doi.org/10.1016/j.aap.2020.105889

work page doi:10.1016/j.aap.2020.105889 2021
[6]

Feng, Z., Ma, X., Zhu, X., Ma, Z., & IEEE. (2018). Analysis of Driver Brake Behavior Under Critical Cut-in Scenarios. Tongji University, 2054–2059. (WOS:000719424500320)

2018
[7]

Hamdar, S., Qin, L., & Talebpour, A. (2016). Weather and road geometry impact on longitudinal driving behavior: Exploratory analysis using an empirically supported acceleration modeling framework. TRANSPORTATION RESEARCH PART C-EMERGING TECHNOLOGIES, 67, 193–213. (WOS:000377734400012). https://doi.org/10.1016/j.trc.2016.01.017

work page doi:10.1016/j.trc.2016.01.017 2016
[8]

Hu, J., Lin, B., Vega, J., & Tsiang, N. (2023). Predictive Models of Driver Deceleration and Acceleration Responses to Lead Vehicle Cutting In and Out. TRANSPORTATION RESEARCH RECORD, 2677(5), 92–102. (WOS:000878362400001). https://doi.org/10.1177/03611981221128277

work page doi:10.1177/03611981221128277 2023
[9]

Jiao, S., Zhang, S., Li, Z., Zhou, B., & Zhao, D. (2020). An Improved Car-Following Speed Model considering Speed of the Lead Vehicle, Vehicle Spacing, and Driver’s Sensitivity to Them. JOURNAL OF ADVANCED TRANSPORTATION, 2020. (WOS:000509536400001). https://doi.org/10.1155/2020/2797420

work page doi:10.1155/2020/2797420 2020
[10]

(Michael), & Zhou, X

Li, L., Jiang, R., He, Z., Chen, X. (Michael), & Zhou, X. (2020). Trajectory data-based traffic flow studies: A revisit. In TRANSPORTATION RESEARCH PART C-EMERGING TECHNOLOGIES (Vol. 114, pp. 225–240). PERGAMON-ELSEVIER SCIENCE LTD. https://doi.org/10.1016/j.trc.2020.02.016

work page doi:10.1016/j.trc.2020.02.016 2020
[11]

Martínez-Vera, E., Bañuelos-Sánchez, P., & Etcheverry, G. (2022). Lane Changing Model from NGSIM Dataset (O. Vergara-Villegas, V. Cruz-Sanchez, J. Sossa-Azuela, J. Carrasco-Ochoa, J. Martinez-Trinidad, & J. Olvera- Lopez, Eds.; Vol. 13264, pp. 25–34). (WOS:000873588100003). https://doi.org/10.1007/978-3-031- 07750-0_3

work page doi:10.1007/978-3-031- 2022
[12]

Mulder, M., Mulder, M., van Paassen, M., Abbink, D., & IEEE. (2005). Effects of lead vehicle speed and separation distance on driver car-following behavior. Delft University of Technology, 399–404. (WOS:000235210800067)

2005
[13]

Nirmale, S., Pinjari, A., & Chakroborty, P. (2024). A two-dimensional, multi-vehicle anticipation, and multi- stimuli based latent class framework to model driver behaviour in heterogeneous, disorderly traffic conditions. TRANSPORTATION RESEARCH PART C-EMERGING TECHNOLOGIES, 160. (WOS:001178257400001). https://doi.org/10.1016/j.trc.2023.104458

work page doi:10.1016/j.trc.2023.104458 2024
[14]

Qian, Z., Li, J., Li, X., Zhang, M., & Wang, H. (2017). Modeling heterogeneous traffic flow: A pragmatic approach. TRANSPORTATION RESEARCH PART B-METHODOLOGICAL, 99, 183–204. (WOS:000401398700009). https://doi.org/10.1016/j.trb.2017.01.011

work page doi:10.1016/j.trb.2017.01.011 2017
[15]

Shi, X., Wong, Y., Li, M., Palanisamy, C., & Chai, C. (2019). A feature learning approach based on XGBoost for driving assessment and risk prediction. ACCIDENT ANALYSIS AND PREVENTION, 129, 170–179. (WOS:000474674100017). https://doi.org/10.1016/j.aap.2019.05.005

work page doi:10.1016/j.aap.2019.05.005 2019
[16]

Svärd, M., Bärgman, J., & Victor, T. (2021). Detection and response to critical lead vehicle deceleration events with peripheral vision: Glance response times are independent of visual eccentricity. ACCIDENT ANALYSIS AND PREVENTION, 150. (WOS:000623628400008). https://doi.org/10.1016/j.aap.2020.105853

work page doi:10.1016/j.aap.2020.105853 2021
[17]

Tavakoli, A., Boker, S., & Heydarian, A. (2023). Driver State Modeling Through Latent Variable State Space Framework in the Wild. IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, 24(2), 1879–1893. (WOS:000890833200001). https://doi.org/10.1109/TITS.2022.3221858

work page doi:10.1109/tits.2022.3221858 2023
[18]

Varotto, S., Jansen, R., Bijleveld, F., & van Nes, N. (2021). Adaptations in driver deceleration behaviour with automatic incident detection: A naturalistic driving study. TRANSPORTATION RESEARCH PART F-TRAFFIC PSYCHOLOGY AND BEHAVIOUR, 78, 164–179. (WOS:000641904800013). https://doi.org/10.1016/j.trf.2021.02.011

work page doi:10.1016/j.trf.2021.02.011 2021
[19]

Wang, Y., Delle Monache, M., & Work, D. (2022). Identifiability of car-following dynamics. PHYSICA D- NONLINEAR PHENOMENA, 430. (WOS:000766758600003). https://doi.org/10.1016/j.physd.2021.133090

work page doi:10.1016/j.physd.2021.133090 2022
[20]

Witt, M., Kompass, K., Wang, L., Kates, R., Mai, M., & Prokop, G. (2019). Driver profiling—Data-based identification of driver behavior dimensions and affecting driver characteristics for multi-agent traffic simulation. TRANSPORTATION RESEARCH PART F-TRAFFIC PSYCHOLOGY AND BEHAVIOUR, 64, 361–376. (WOS:000482515300028). https://doi.org/10.1016/j.trf.2019.05.007

work page doi:10.1016/j.trf.2019.05.007 2019
[21]

Xue, Q., Ouyang, X., Zhao, Y., & Guo, W. (2022). Effect of Situation Kinematics on Drivers’ Rear-End Collision Avoidance Behaviour-A Combined Effect of Visual Looming, Speed, and Distance Analysis. SUSTAINABILITY, 14(22). (WOS:000887637600001). https://doi.org/10.3390/su142215103

work page doi:10.3390/su142215103 2022
[22]

Zheng, P., & McDonald, M. (2001). Identifying best predictors for car following behaviour from empirical data. In N. Giambiasi & C. Frydman (Eds.), University of Southampton (pp. 158–165). (WOS:000174740600028)

2001