Driver Behavior Estimation at Signalized Intersections Using a Physics-Constrained Decision-Conditioned Autoregressive Transformer
Mohammad Khoshkdahan, Pavel Laskov, Alexey Vinel
cs.LG · cs.AI · cs.CY · eess.SY
Quick verdict
Novel RTK driver dataset + two-stage transformer for yellow-onset stop/go and braking; validation-split metrics and thin baselines cap confidence.
Summary
The paper introduces a real-world dataset of 449 vehicle approach runs at a signalized intersection, collected from 30 licensed drivers using RTK-GNSS (centimeter-level, 10 Hz), continuous heart-rate monitoring, and five-level post-stop comfort ratings; a portable traffic light was triggered from vehicle position with latency calibration (142 ms; onset error below 30 cm). A Bayesian hierarchical logistic model identifies required deceleration a_req as the dominant stop/go predictor (average marginal effect +33.7 percentage points), and heteroscedastic Gaussian Bayes boundaries yield five comfort ranges (peak deceleration 4.07-9.11 m/s2). The proposed two-stage framework uses an MLP for the stop/go decision, then a decision-conditioned autoregressive Transformer generates the longitudinal acceleration trajectory under physics constraints (terminal gate, 15 m/s3 jerk limit, actuator clamps). On a stratified 60/40 split it reports 0.49 m/s2 acceleration MAE, 0.62 m distance MAE, 95.5% stop-band success, and 84.5% driver-comfort matching; the compact (speed, a_req) Stage-1 configuration reaches 92.4% accuracy, dropping to 87.9% under a participant-disjoint split. Main limitations: trajectory-stage results are not reported under participant-disjoint splits, headline metrics come from a validation split with no separate test set, learned sequence baselines are absent, the trivial constant-deceleration baseline attains higher stop-band success (100%), and no ethics approval for the human-subjects study is stated.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- This is the first real-world dataset combining centimeter-level RTK vehicle motion with synchronized driver heart rate and multi-level stopping-comfort ratings at signalized intersections (449 runs, 30 drivers).
- Required deceleration a_req is the dominant single predictor of the stop/go decision (AME +33.7 percentage points; 50.8% relative contribution, Table I).
- The two-stage DCAR Transformer achieves 0.49 m/s2 acceleration MAE and 0.62 m distance MAE on a stratified split, beating all tested baselines on tracking errors (Table II: 0.499/0.628).
- Driver stopping-comfort level can be estimated from a single yellow-onset snapshot, with generated profiles matching ground-truth comfort in 84.5% of stop cases.
Strengths
Novel high-precision multimodal dataset
449 runs / 30 drivers combining cm-level RTK-GNSS at 10 Hz, synchronized heart rate, and 5-level stopping-comfort ratings (Sec. III); the vehicle-triggered signal chain is calibrated for a measured 142 ms latency with <30 cm residual onset error (Sec. III-C) — a genuinely careful calibration treatment absent from camera-based predecessors.
Decision analysis is statistically principled
Bayesian hierarchical logistic model with participant random intercepts (Eq. 5); Table I reports AMEs rather than raw coefficients, quantifying a_req dominance (+33.7 pp, 50.8% relative contribution), and Fig. 4 shows the decision boundary shifting with distance rather than following a fixed deceleration threshold.
Extensive feature ablations
Physics constraints are empirically grounded and internalized
Open data and code
Weaknesses
Stage 2 never evaluated participant-disjoint
Validation split doubles as test
Thin baseline set with an awkward inversion
Human-subjects governance unstated
Some demographic/HR claims outrun the statistics
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper makes three coupled contributions: (1) a field dataset of 449 approach runs (392 stop/go decision runs plus forced-stop runs; 185,109 GNSS samples) from 30 licensed drivers at a purpose-instrumented test intersection in Germany; (2) a statistical characterization of the yellow-onset stop/go decision and of braking comfort; (3) a two-stage decision-conditioned autoregressive (DCAR) Transformer that, from a single yellow-onset snapshot, predicts the maneuver and generates the full longitudinal acceleration trajectory. Headline claims: acceleration MAE 0.49 m/s2 and distance MAE 0.62 m on a stratified split (Table II row 'DCAR Transformer (ours)': 0.499/0.628), 95.5% stop-band success, 84.5% ground-truth comfort matching, and required deceleration a_req as the dominant decision predictor (AME +33.7 pp, 50.8% relative contribution, Table I). A 'first real-world dataset' priority claim is asserted for the combination of RTK motion, physiology, and comfort ratings; adjacent datasets (SIND) are dismissed in Sec. II without a systematic comparison table.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does the paper's model predict?
From a single observation at yellow-light onset (approach speed, distance to the stop line, derived kinematics, and demographics), it first predicts whether the driver will stop or go, then generates the complete longitudinal acceleration/deceleration trajectory until the vehicle stops or crosses the intersection, under physics constraints such as a jerk limit.
How accurate is the model?
On a 60/40 stratified split the authors report 0.49 m/s2 acceleration MAE, 0.62 m distance MAE, 95.5% of predicted stops landing in the stop band, and 84.5% agreement with the driver's comfort rating. The stop/go decision reaches 92.4% accuracy, dropping to 87.9% when evaluated on drivers never seen in training.
What data was collected and is it available?
449 approach runs from 30 licensed drivers at a test traffic light in Germany over 8 days, recorded with centimeter-level RTK-GNSS at 10 Hz, a wearable heart-rate sensor, and 1-5 comfort ratings after each stop. The paper states that the dataset and source code are publicly available on GitHub (driver-2s-transformer).
What are the five comfort ranges reported?
By modeling each comfort rating class as a Gaussian distribution over peak braking deceleration and applying a Bayes-optimal boundary rule, the authors derive transition thresholds of 4.07, 5.15, 7.54, and 9.11 m/s2 separating ratings 5 through 1, i.e., from very smooth stopping to very stressful braking.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are available only as extracted text fragments; visual claims (e.g., trajectory shapes in Fig. 7, decision-surface geometry in Fig. 4) could not be inspected directly.
- The stated public GitHub repository was not inspected; code/data availability is taken from the authors' claim.
- Review is based on arXiv v1 (13 Sep 2026); no peer-review history or supplementary material was available.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.