Automated Brake Onset Detection in Naturalistic Driving Data

Gustav Markkula; Johan Engstr\"om; Shu-Yuan Liu

REVIEW 5 major objections 5 minor 2 references

Automated Brake Onset Detection in Naturalistic Driving Data

T0 review · 5 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 By fitting a two-piece straight-line model to longitudinal acceleration, this paper aims to show that brake onset in traffic conflicts can be estimated automatically, with 91.1% of estimates within 0.5 seconds of human annotation.

desk verdict Solid, modest method paper for estimating brake onset from kinematics; the accuracy numbers need an inter-rater reliability check before they mean what they claim. read the letter →

arxiv 2507.17943 v2 pith:KB7AU47R submitted 2025-07-23 cs.HC cs.RO

Shu-Yuan Liu , Johan Engström , Gustav Markkula This is my paper · ORCID

classification cs.HCcs.RO

keywords brakeonsetdetectionpiecewiselinearmodeltrafficconflictsnaturalisticdrivingdataautomatedsystemsresponsetiminglongitudinalaccelerationconfidencemetric

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

The paper sets out to show that the moment a driver or automated vehicle starts an evasive brake can be read directly from longitudinal acceleration, without needing pedal, brake-light, or steering signals. It does this with a two-piece piecewise linear model: acceleration holds at a constant level, then drops along a straight line, and the time the line begins to drop is the estimated brake onset. Validated against manual annotation on 190 real traffic conflicts, 91.1% of estimated onsets were within 0.5 seconds and 84.2% within 0.3 seconds, with the $R^2$ fit quality serving as a confidence score that separates good from poor estimates at AUC 0.71. This matters because automated-driving logs and dense naturalistic data often contain kinematics of surrounding road users but no control signals, and response timing is central to benchmarking and evaluating collision avoidance.

What carries the argument

The central object is a two-piece piecewise linear model of longitudinal acceleration, with parameters $a_0$ (initial constant acceleration), $t_B$ (brake onset), and $j_B$ (constant jerk during braking). A grid search over each parameter is run for every event, the fit window runs from one second before the manually annotated stimulus onset $T_1$ to the minimum-acceleration time $T_{a_{\min}}$ (or to 0.2 s before a crash), and the candidate with highest $R^2$ is selected. The model returns $t_B$ as the estimated brake onset and the winning $R^2$ as a confidence score; the score is what lets users automatically keep reliable estimates and route uncertain ones to manual review.

What would settle it

Compare the model's $t_B$ against brake-pedal activation timestamps in a naturalistic dataset where pedal signals exist; if the agreement rate at the 0.5 s tolerance drops below the reported 91.1%, the kinematic-only claim is not general. Alternatively, rerun the pipeline with an automatic conflict-detection anchor, such as a minimum time-to-collision, in place of the manual $T_1$ and measure how many estimates remain within tolerance.

Watch

Extended reading notes

Core claim

The central claim is that brake onset in collision-avoidance events is identifiable from longitudinal acceleration alone, using a deliberately simple kinematic model: a constant initial acceleration $a_0$, a switch time $t_B$, and a constant jerk $j_B$ describing the deceleration ramp. On 190 naturalistic crash and near-crash events where both the model and manual annotation produced a value, the model's $t_B$ was within 0.5 s of the manual annotation in 91.1% of events and within 0.3 s in 84.2%, with estimates tending to run slightly early. The paper also shows that the $R^2$ of the best-fitting model is a usable confidence metric: it correlates negatively with absolute error, and a classifier based on $R^2$ separates small-difference from large-difference estimates with an AUC of 0.71. The deviations that do exceed 0.5 s are traced to pre-existing slowing, stepwise or multi-ramp braking, noisy signals, and contexts where human annotators also need video to decide.

Load-bearing premise

The method presupposes that a traffic conflict has already been identified and that a stimulus-onset time $T_1$ is available to anchor the fit window, and the one-second and 0.2-second offsets in that window were chosen after observing the data.

Editorial extensions

If this is right

Brake response timing can be extracted at scale from Level 4 ADS logs and naturalistic data where only longitudinal kinematics of surrounding road users are available.
An $R^2$ threshold gives users a tunable triage: higher thresholds admit fewer false positives at the cost of excluding some good estimates, letting them decide how much manual review to do.
Because the window is anchored only by an estimate of when a conflict might occur, the method can be reused with alternative anchors such as time-to-collision or post-encroachment time, not only a manual stimulus onset.
The same algorithm applies across the road-user types and conflict configurations studied here, including non-car users, with similar error distributions, though foot-braking behavior remains a caution case.

Reading between the lines

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

A natural next test is to fully automate the conflict anchor: replace the manual $T_1$ with a surrogate-measure event time and measure the drop in accuracy; the paper only notes this is possible without testing it.
For stepwise and multi-ramp braking, which the paper identifies as the largest failure class, a model with two brake ramps or a curvature-based knee detector might outperform the single-ramp model on the same data.
The same $R^2$-confidence idea could serve as a screening tool for lateral swerve onset, even though the paper reports that its own lateral-fit attempts failed, because the confidence score would let analysts know when the fit should not be trusted.
If the method is applied to vulnerable road users who brake with their feet, additional smoothing or a minimum-deceleration filter will likely be required; the paper itself suggests caution and reports that sample sizes were too small to settle this.

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

Here's the quick take: this is a useful, clearly written method paper that does what it says—estimates brake onset from longitudinal acceleration alone—but its headline accuracy numbers are anchored to a single human annotator, so they are not yet as strong as they look.

The genuinely new part is the two-piece piecewise linear model, a simplification of Markkula et al.'s three-piece PLM with the final acceleration parameter removed, per-event grid search, and an R-squared-based confidence metric. Applied to 219 naturalistic conflicts including VRUs and eight scenario types, it reports 91.1% of 190 estimated onsets within 0.5 s of manual annotation, and an AUC of 0.71 for using R-squared to flag bad estimates. The algorithm is simple, efficient, and the authors are unusually honest: they analyze the 17 outliers, discuss the single-brake-ramp assumption, acknowledge that foot-braking VRUs are a problem, and note that practical application presupposes some conflict-detection anchor. That is real credit.

The soft spot is exactly what the stress-test says: there is no second annotator and no inter-rater reliability statistic. All comparisons use the same person's manual labels. If human-human agreement on brake onset from acceleration plots plus video is in the same 0.3-0.5 s range as the model-human deviations, then the accuracy claim is partly measuring annotation noise. The T1 anchor defining the fit window is also manually annotated, and the 1 s / 0.2 s offsets were chosen by looking at the data. That doesn't make the method circular—the tB estimate is not fit to the manual onset labels—but it does mean the reported performance is conditional on the quality of the T1 and on the window offsets. The 0.3 s and 0.5 s thresholds are arbitrary, and the AUC is in-sample. A simple baseline (e.g., a fixed deceleration threshold or a two-point jerk detector) would have made the contribution clearer.

None of this sinks the paper. As a method contribution for ADS safety evaluation, it fills a real gap: estimating response onset from kinematics of surrounding road users without pedal signals. The right audience is traffic safety researchers and ADS evaluation engineers. It deserves a serious referee. For the revision, I'd ask for a multi-annotator reliability check on at least a subset, and a comparison against a simple threshold-based onset detector. With those, the empirical claims would be solid.

Referee Report

5 major / 5 minor

Summary. The paper proposes a method for estimating the onset of evasive braking from longitudinal acceleration time series in traffic conflicts. The core algorithm fits a two-piece piecewise linear model (constant acceleration followed by a linear deceleration ramp) to a window anchored at a manually annotated stimulus onset T1; the breakpoint time is the estimated brake onset, and the R^2 of the best fit is proposed as a confidence metric. The method is evaluated on 219 naturalistic conflicts involving Waymo rider-only vehicles and other road users, of which 190 produce numeric deviations from manual annotation. The authors report 91.1% of these 190 deviations within 0.5 s and 84.2% within 0.3 s, and an AUC of 0.71 for an R^2-based classifier of accurate versus inaccurate estimates. The paper also describes a manual annotation procedure, analyzes failure cases (crashes, non-car road users, no-response events), and discusses the limitations of the single-ramp model, the exclusion of steering, and the need for a conflict anchor.

Significance. Measuring response onset in large ADS logs without access to vehicle control signals is a real bottleneck for benchmarking and safety evaluation. If the reported accuracy were established, this simple, interpretable model with its R^2-based selection rule would be a useful, low-cost tool for a range of applications in naturalistic driving analysis and simulator-based ADS evaluation. The paper's strengths include the use of naturalistic data spanning multiple conflict types and road-user types, a transparent two-piece model with a configurable grid search, a detailed failure analysis that goes beyond aggregate percentages, and an unusually candid Discussion of the method's assumptions and boundaries. However, the validation is anchored to a single human annotator, and the fit window depends on manual T1 and data-selected offsets. The headline accuracy therefore has not yet been separated from annotation uncertainty or shown to transfer to a fully automatic conflict-detection pipeline. These issues are substantive but addressable, so the contribution is promising rather than conclusive.

major comments (5)

[Manual ground truth annotation] The validation is anchored entirely to one human annotator. The 'Manual ground truth annotation' section describes the annotation of brake onset from acceleration plots plus video, but no inter-annotator reliability is reported, and the same annotator also provides the T1 stimulus onset that defines the model fit window (Methods, 'Brake onset estimation with piecewise linear model'). Consequently, the reported 91.1% within 0.5 s and 84.2% within 0.3 s (Results, Figure 5) could reflect human annotation uncertainty rather than algorithmic skill, especially since the paper itself notes that some events are challenging for humans (e.g., the -4.4 s outlier discussion). Please add at least a second annotator on a random subset and report human-human agreement (e.g., proportion of differences within 0.3/0.5 s or an appropriate interval agreement statistic), and show that the model-human differences are no larger than human-human differences.
[Brake onset estimation with piecewise linear model] The model fit window is defined as [T1-1 s, T_a_min], where T1 is manually annotated, and the 1 s and 0.2 s offsets are 'selected based on observing the data.' The reported validation is therefore conditional on manual T1 and on data-dependent window offsets. The Discussion acknowledges that practical application presupposes a way to identify conflicts and suggests surrogate anchors such as minimum time-to-collision. Please evaluate the sensitivity of the t_B estimates to the window offsets (e.g., systematically vary the 1 s and 0.2 s values) and, if feasible, test the pipeline with an automatically derived anchor in place of T1. Without such an analysis, the demonstrated performance does not yet extend to the fully automatic setting that motivates the paper.
[Results, Overall comparison] The headline accuracy figures are computed on the 190 events with numeric deviations, after excluding 26 events with missing deviations (no model output, no braking, or no response) and 3 events with incomplete time series. Across all 219 sampled conflicts, this corresponds to at most 79% of events within 0.5 s if the excluded events are treated as failures. The separate analysis in 'Events with missing deviation values' is informative, but the main claim should be stated with the denominator made explicit, and the paper should recommend how practitioners should handle events where the algorithm returns no numeric onset (e.g., always send them to manual review, or filter by an a_min threshold). Please add a coverage statement to the Results and Abstract.
[Results, Model fit in different road user groups] The claim that the method generalizes to different road user types and scenario types is based on very small samples (e.g., 5 bicycles, 1 micromobility device, 2 pedestrians in Table 1; 15 VRU scenarios in Table 2) and on visual inspection of the distributions in Figures 9-11, with no inferential statistics behind the 'similar distributions' claim. Please provide per-group summary statistics (median, IQR, and within-threshold proportions with confidence intervals) or explicitly limit the generalizability claim to passenger cars and frame the VRU results as exploratory. This is important because the paper's motivation is broad applicability across road user types.
[Results, Confidence metrics for results selection] The AUC of 0.71 for the R^2 confidence metric is computed in-sample on the same 190 events that were used to select the window offsets and the 0.3 s tolerance. Because the R^2 values come from the same model that produced the t_B estimates, and no held-out or cross-validated evaluation is reported, the ROC curve in Figure 14 may overstate the practical utility of R^2 as a selection tool on new data. Please provide a cross-validated or split-sample estimate of the AUC, and consider also reporting precision-recall curves since the classes are imbalanced (roughly 84% small-difference events).

minor comments (5)

[Abstract and Results, Overall comparison] Please qualify the 91.1% and 84.2% figures by explicitly stating the denominator (190 numeric-deviation events) and noting that 29 of 219 events were excluded; consider reporting the coverage rate in the Abstract.
[Table 3] The row totals are inconsistent: 1 + 14 + 11 = 26, but the passenger-car total is listed as 25. The note that the single no-model-output event was also a no-braking event means the categories overlap; please clarify whether the categories are mutually exclusive or use a footnote to explain the double counting.
[Figure 5] Please add units (seconds) to the x-axis label and describe the bin width in the caption to make the histogram reproducible.
[Methods, Definitions and Discussion] The paper says the model 'in principle only requires a single input signal: longitudinal acceleration' but the current implementation also uses jerk (from finite differences) to set grid-search ranges and uses manually annotated T1; please reconcile this description with the actual inputs.
[Results and References] Please report the regression p-value as p < 0.001 rather than 'p-value=0.00', and run a citation consistency check: for example, Shapiro and Wilk (1965) appears in the reference list but is not cited in the text, and several inline citations use inconsistent formatting such as 'Markkula et. al'.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: brake onset is obtained by fitting a piecewise linear model to acceleration traces, not by fitting to the manual ground-truth labels; the residual concerns are the manual T1 anchor, in-sample window choices, and single-annotator validation.

full rationale

The derivation chain is: (1) a human annotator defines brake onset from acceleration time-series plots and event video; (2) a two-piece piecewise linear model is fitted to the same acceleration time series within a window anchored at manually annotated stimulus onset T1, selecting parameters by maximum R^2 with respect to the acceleration signal; (3) the fitted breakpoint t_B is compared with the manual onset; and (4) R^2 is evaluated as a confidence classifier against the resulting absolute deviations. At no point do the manual brake-onset labels enter the model fitting objective: the grid search maximizes R^2 against acceleration alone, so t_B is not a relabeling of the ground truth and the 91.1% within 0.5 s figure is an honest, though in-sample, comparison. The main dependency is the manually annotated T1, which defines the model fit window; the paper itself states that practical application presupposes a way to identify conflicts and that T1 came from Engström et al. (2024). The 1 s and 0.2 s window offsets were also selected by observing the data, and the ground truth comes from a single annotator without inter-rater reliability. These are genuine limitations affecting generalizability and validation strength, but they are not circular reductions: no quoted equation defines t_B in terms of the manual onset, and no self-citation is used to forbid alternative models or to supply the predicted values. The self-citations to Engström et al. (2024) and Markkula et al. (2016) provide the annotation convention and the PLM modeling ansatz, respectively, but the central estimate remains independent of those citations' conclusions. Overall, there is no self-definitional, fitted-input-renamed-as-prediction, or uniqueness-imported circularity; the paper is best characterized as having minor self-citation and validation-dependency concerns rather than a circular derivation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim relies on a handful of hand-set constants, the availability of manually annotated T1, and the single-ramp approximation of braking. The model parameters a0, tB, and jB are per-event fitted outputs rather than hidden inputs. No new physical entities are introduced.

free parameters (4)

Fit window offsets (t_start, crash cutoff) = 1 s before T1; 0.2 s before crash
The 1 s and 0.2 s values were selected based on observing the data (Methods, Model fit window), and they determine which acceleration segment the PLM sees.
Difference thresholds for good estimation = 0.3 s; upper bound 0.5 s
The 0.3 s threshold was chosen as a practical error tolerance with no prior standard; 0.5 s is called a reasonable upper bound based on empirical observations (Results, Confidence metrics).
Grid search ranges and steps = a0 in [amax-1, amax+1] step 0.1; tB in [tstart, tend] step 0.1; jB in [jmin-5, 0] step 0.2
Grid bounds are customized per event from observed kinematics to improve efficiency; they constrain the fitted tB.
Optional amin filter threshold = -0.3 or -0.4 m/s^2
Explored post hoc to flag no-braking or no-response events; not adopted in the main algorithm but illustrates an additional hand-set threshold.

assumptions (5)

domain assumption The manually annotated brake onset is the correct ground truth
All validation compares the algorithm to single manual annotations of the same acceleration plots; no inter-annotator reliability is reported (Methods, Manual ground truth annotation).
domain assumption A stimulus onset T1 is available to define the fit window
The fit window is [T1-1 s, T_a_min] for near-crashes and [T1-1 s, 0.2 s before crash] for crashes; without T1 the algorithm has no window (Methods).
domain assumption Evasive braking can be approximated as a single brake ramp after a constant acceleration phase
The two-piece PLM has constant a0 followed by a linear decrease; stepwise braking, prior slowing, and foot-braking VRUs violate this, as acknowledged in the Discussion.
domain assumption Longitudinal acceleration time series from Waymo perception are accurate enough for onset detection
The algorithm uses unfiltered acceleration from logs for responders, including non-ego road users; no sensor error model is given.
standard math Standard grid search and R-squared model selection locate the true breakpoint
The method assumes the highest-R-squared piecewise linear fit over the grid gives the brake onset; there is no proof of identifiability for tB under noise.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Automated Brake Onset Detection in Naturalistic Driving Data." pith.science (2026). https://pith.science/paper/KB7AU47R

@misc{pith2026250717943,
  author       = {Pith},
  title        = {Pith review of: Automated Brake Onset Detection in Naturalistic Driving Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KB7AU47R}},
  note         = {Machine review of arXiv:2507.17943}
}

read the original abstract

Response timing measures play a crucial role in the assessment of automated driving systems (ADS) in collision avoidance scenarios, including but not limited to establishing human benchmarks and comparing ADS to human driver response performance. For example, measuring the response time (of a human driver or ADS) to a conflict requires the determination of a stimulus onset and a response onset. In existing studies, response onset relies on manual annotation or vehicle control signals such as accelerator and brake pedal movements. These methods are not applicable when analyzing large scale data where vehicle control signals are not available. This holds in particular for the rapidly expanding sets of ADS log data where the behavior of surrounding road users is observed via onboard sensors. To advance evaluation techniques for ADS and enable measuring response timing when vehicle control signals are not available, we developed a simple and efficient algorithm, based on a piecewise linear acceleration model, to automatically estimate brake onset that can be applied to any type of driving data that includes vehicle longitudinal time series data. We also proposed a manual annotation method to identify brake onset and used it as ground truth for validation. R^2 was used as a confidence metric to measure the accuracy of the algorithm, and its classification performance was analyzed using naturalistic collision avoidance data of both ADS and humans, where our method was validated against human manual annotation. Although our algorithm is subject to certain limitations, it is efficient, generalizable, applicable to any road user and scenario types, and is highly configurable.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 linked inside Pith

[1]

How long does it take to stop?

Dingus, T. A., Klauer, S. G., Neale, V. L., Petersen, A., Lee, S. E., Sudweeks, J., ... & Knipling, R. R. (2006). The 100-car naturalistic driving study, Phase II-results of the 100-car field experiment (No. DOT-HS-810-593). United States. Department of Transportation. National Highway Traffic Safety Administration. Dinparastdjadid, A., Supeene, I., & Eng...

arXiv 2006
[2025]

arXiv preprint arXiv:2506.02215

Active inference as a unified model of collision avoidance behavior in human drivers. arXiv preprint arXiv:2506.02215. SAE International (2015) Operational Definitions of Driving Performance Measures and Statistics. SAE International (2021) Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles. 22 Terranova, P...

arXiv 2015

Pith tools

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

[1] [1]

How long does it take to stop?

Dingus, T. A., Klauer, S. G., Neale, V. L., Petersen, A., Lee, S. E., Sudweeks, J., ... & Knipling, R. R. (2006). The 100-car naturalistic driving study, Phase II-results of the 100-car field experiment (No. DOT-HS-810-593). United States. Department of Transportation. National Highway Traffic Safety Administration. Dinparastdjadid, A., Supeene, I., & Eng...

arXiv 2006

[2] [2025]

arXiv preprint arXiv:2506.02215

Active inference as a unified model of collision avoidance behavior in human drivers. arXiv preprint arXiv:2506.02215. SAE International (2015) Operational Definitions of Driving Performance Measures and Statistics. SAE International (2021) Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles. 22 Terranova, P...

arXiv 2015