Source-linked AI summary
Counterfactual Closing-Acceleration Risk: An Anticipatory Surrogate Safety Measure for the Blind Region of Car-Following
Eni Solomon Laughter
TL;DR
Proximity-based surrogate measures leave many car-following states unscored because they assume invariant motion and require the follower to be faster. This paper introduces CCAR, which counterfactually projects leader braking using the follower’s current closing acceleration. On expressway trajectories, CCAR provides graded risk across the blind region, while simulations support closing acceleration as a collision-related precursor.
Problem
Proximity surrogate measures cannot represent latent risk in states where the follower is not yet faster, despite differences in follower acceleration at identical gap and relative speed.
Method
CCAR estimates exposure by projecting a hypothetical leader-braking event against the follower’s current motion and closing acceleration, using the minimum projected gap.
Results
51.2% of following frames lie in the blind region, where CCAR is non-zero in 98.8% and produces a counterfactual collision in 7.6%.
Takeaways & Limitations
CCAR supplies an anticipatory risk reading where proximity measures are silent and supports closing acceleration as information beyond gap and speed.
Takeaways & Limitations
CCAR is conditional rather than predictive: its parameters are assumptions, and trajectory data lack crashes for outcome validation.
Abstract
from arXiv · showhide
Surrogate safety measures allow road-safety assessment from trajectory data in which crashes are absent, yet the dominant proximity measures - time-to-collision (TTC) and its variants - assume invariant motion and are undefined whenever the following vehicle is not yet faster than its leader, so a large fraction of car-following carries no risk reading at all. This paper introduces Counterfactual Closing-Acceleration Risk (CCAR), an anticipatory surrogate measure that scores how exposed a follower is to a rear-end conflict if its leader were to brake, conditioned on the follower's current gap-closing acceleration. CCAR is evaluated on 745,540 expressway car-following frames from the SQM-W-1 trajectory dataset. Conventional measures leave 51% of frames unscored because the follower is not yet faster; across this blind region CCAR returns a graded, non-trivial risk in 98.8% of frames. CCAR is not redundant with existing measures (Spearman correlation 0.54 with modified TTC; top-decile risk-set overlap 0.19), and at fixed gap and speed its risk rises monotonically with closing acceleration. A controlled simulation with a reactive Intelligent Driver Model follower and a scripted leader brake confirms that, holding gap and brake fixed, the actual collision rate rises with closing acceleration, supporting the proposed precursor mechanism. Parameter sensitivity, limitations, and implications for forward-collision-warning systems are discussed.
1. Introduction
Surrogate safety measures enable proactive road-safety assessment from trajectory data, but dominant proximity measures miss latent risk when follower acceleration changes. CCAR addresses this gap by assigning anticipatory risk to car-following states based on a hypothetical leader-braking event and current gap-closing acceleration.
- Motivation: Surrogate safety measures quantify near-collision risk from trajectory data without requiring crashes to occur.High-resolution trajectories provide position, speed, and acceleration for microscopic interaction analysis.
- Motivation: Car-following is the proximate setting of rear-end risk, and follower acceleration is a measurable, behaviorally meaningful descriptor of following state.Driver heterogeneity means otherwise similar situations can involve different gaps, accelerations, and braking responses.
- Gap in existing measures: TTC is defined only when the follower is already faster than its leader, assigning infinite TTC to slower or matched states even when the follower accelerates toward the gap.Related measures such as DRAC and MTTC address some limitations but do not read risk from this specific precursor.
- Proposed measure: CCAR assigns scalar exposure to a rear-end conflict if the leader brakes now, conditioning a short counterfactual projection on the follower’s current gap-closing acceleration.Its mechanism is summarized by the minimum projected gap.
- Study aims: The study evaluates whether CCAR identifies TTC-safe but risky states and whether closing acceleration adds risk information beyond gap and speed.These questions are examined using real expressway trajectory data and controlled simulation.
2. Problem Statement
Proximity measures are systematically blind to latent rear-end risk in states where the follower is not yet faster than the leader. The paper therefore requires a measure that incorporates follower acceleration and remains interpretable without crash outcomes.
- Blind state: When a follower is no faster than its leader but is accelerating, TTC is infinite and DRAC is zero even though leader braking and reaction lag may drive the gap shut.Under constant-velocity assumptions, the relative speed is non-positive and the projected gap never decreases.
- Blind state: Proximity measures cannot distinguish identical-gap, identical-relative-speed states when follower acceleration differs, despite different latent risk.One follower may be accelerating into the gap while another is easing off.
- Requirements: A resolving measure must be informative before the follower becomes faster, depend on follower acceleration, and remain interpretable without crash outcomes.CCAR is constructed to meet these three requirements.
3. Related Work
Related work includes proximity, field-based, deceleration- and reaction-time-based, extreme-value, and behavioral approaches to surrogate safety. CCAR differs by conditioning a counterfactual braking exposure on follower closing acceleration.
- Proximity measures: TTC, DRAC, and MTTC dominate longitudinal surrogate-safety practice but rely on closing motion or thresholds, with MTTC still keyed to current relative motion.These measures are simple and interpretable but do not directly specify a braking event.
- Field-based measures: Field and potential-based measures provide continuous risk surfaces and can return non-zero risk where proximity measures cannot.Prior work explicitly addresses motion-state invariance and supports risk estimation during otherwise unscored windows.
- Braking and reaction-time measures: Deceleration- and reaction-time-based measures account for braking capability, required deceleration, or driver reaction time when estimating rear-end danger.CCAR shares this family’s attention to reaction time and braking assumptions while adding closing-acceleration conditioning.
- Crash-frequency inference: Extreme value theory extrapolates crash frequency from severe observed conflicts, whereas CCAR quantifies latent exposure before a conflict occurs.The two approaches address sparse crash data from opposite directions.
- Behavioral motivation: Behavioral studies motivate CCAR by documenting acceleration to claim gaps, anticipation during lane changes, and acceleration-related conflict contributions.These findings connect following risk to behavior beyond instantaneous proximity.
4. The CCAR Measure
CCAR projects a hypothetical leader brake against the follower’s current motion, including a reaction window, then converts the minimum projected gap into a normalized risk score. Its distinguishing dynamics arise because leader braking and follower closing acceleration jointly reduce the gap during reaction.
- Counterfactual construction: CCAR asks how close vehicles would come if the leader braked now while the follower continued current motion through a reaction window before braking.The measure is a counterfactual exposure score rather than a direct crash probability.
- Inputs and assumptions: The projection starts from leader speed, follower speed, gap, and nonnegative closing acceleration, plus assumed leader braking, reaction-time, and follower-braking parameters.Negative measured follower acceleration is set to zero so the score reflects genuine closing.
- Follower response: The follower maintains current acceleration during the reaction window and then brakes according to its modeled braking capability.The construction uses the intermediate reaction-window speed and displacement before the braking phase.
- Projected gap: The projected gap equals initial gap plus leader displacement minus follower displacement over the counterfactual trajectory.This decomposition directly separates initial spacing, leader motion, and follower motion.
- Discriminator: During the reaction phase, gap curvature contains the additive term −(b_L + α), making leader braking and follower closing acceleration the dynamics that distinguish CCAR from TTC.The discriminator set consists of states where TTC is undefined or above a safe threshold while CCAR reports risk.
- Risk read-outs: The minimum projected gap is converted into a normalized risk score, with counterfactual collision defined by a non-positive minimum gap.Collision severity is read from closing speed when the projected gap first reaches zero.
- Parameter assumptions: The three braking and reaction parameters are modeling assumptions swept across 27 combinations rather than treated as measured constants.The paper reports no conclusion that depends on one chosen parameter setting.
5. Data and Methods
The study extracts high-resolution car-following frames from the SQM-W-1 expressway dataset, audits kinematic consistency, and computes CCAR alongside conventional SSM baselines across parameter settings.
- Dataset: 822,712 vehicle-frame records cover 1,041 vehicles across ten lanes at 24 Hz, with positions, speeds, accelerations, identities, and gap distances.The data come from a dual-carriageway expressway segment containing a continuous merge-influence zone.
- Data quality: 0.0417 s is the regular sampling interval used to audit consistency among stored position, speed, and acceleration fields.The acceleration-driven measure therefore depends on verified kinematic timing and derivative relationships.
- Data quality: 0.052 m/s is the position-derived speed RMSE, while speed-derived acceleration differs from stored acceleration by 0.134 m/s².Implausible accelerations beyond ±8 m/s² account for 0.046% of frames, and stored acceleration is smoother than the derived value.
- Frame selection: 745,540 car-following frames remain after excluding missing leaders, non-positive gaps, and implausible accelerations.This represents 90.6% of the raw records; TTC, DRAC, and MTTC are computed per frame as baselines.
- CCAR evaluation: 27 parameter settings combine three values each for leader braking, reaction time, and follower braking capability.The central CCAR setting uses b_L = 6 m/s², t_r = 1.0 s, and b_F = 6 m/s².
6. Results
CCAR assigns graded counterfactual risk in states that proximity measures classify as safe, while adding information beyond conventional baselines and varying across traffic locations and assumptions.
- Existence and prevalence: 14.7% of all following frames yield a counterfactual collision at the central setting, with mean normalized risk of 0.600.Among TTC-safe frames, counterfactual-collision proportions are 14.4% at 1.25 s, 14.3% at 1.5 s, and 13.6% at 4 s; median and 90th-percentile closing speeds are 5.4 and 8.3 m/s.
- Existence and prevalence: 51.2% of following frames lie in the blind region where TTC is infinite and DRAC is zero, yet CCAR is non-zero in 98.8% of those frames.Blind-region CCAR has median 0.485, 90th percentile 0.922, and 7.6% counterfactual collisions.
- Incremental information: 0.39, 0.25, and 0.54 are CCAR’s Spearman correlations with TTC, DRAC, and MTTC across closing frames, respectively.The CCAR–MTTC top-decile risk-set overlap is 0.21, so the highest-risk sets are substantially different.
- Incremental information: 0.54 to 0.91 is the example CCAR risk increase across closing-acceleration bands at fixed gap and speed.The rise occurs in most populated cells but is not strictly monotonic in every cell, with some declines at the highest acceleration band.
- Spatial characterization: 19.2%, 23.1%, and 18.9% are collision rates in lanes 3, 4, and 5, exceeding the 11–13% rates in full-length inner and opposing lanes.Closing-acceleration-driven collisions also concentrate in these partial and transition lanes at 5.1%, 3.7%, and 3.4%.
- Robustness across assumptions: 0.3% to 66.4% is the discriminator range across 27 parameter settings, with a median near 15% of TTC-safe frames.The discriminator rises with assumed leader braking and reaction time and falls with follower braking capability, remaining non-zero at every setting.
- Robustness across assumptions: 0.36–0.62 is the CCAR–MTTC Spearman-correlation range across settings, while conditional risk rises with closing acceleration except at one extreme corner.Closing-acceleration-driven spatial concentration is more stable than mean-risk spatial ranking, with median rank correlations of 0.71 and 0.42, respectively.
7. Simulation Validation
Controlled simulation tested whether closing acceleration predicts collisions under a scripted leader brake. Holding gap and braking fixed, collision rates increased with closing acceleration, while CCAR’s risk ranking tracked realized danger but underestimated its level under the tested follower behavior.
- Simulation design: 700 scenarios produced 4,200 leader-brake forks using an IDM follower and a scripted-braking leader across sampled driver parameters.The simulation included timid-to-aggressive IDM settings and realistic following states.
- Causal effect of closing acceleration: At 6 m/s² leader braking and 12–20 m gaps, contact rate increased from 0.22 to 0.60 across closing-acceleration bands.The same monotone rise appeared for 20–40 m gaps.
- Causal effect of closing acceleration: Holding gap and leader braking fixed, actual collision rate rose with closing acceleration, providing controlled evidence for its causal association with collisions under leader braking.This complements the observational association reported elsewhere in the study.
- Calibration against a reactive follower: CCAR’s predicted minimum gap correlated 0.59 with realized minimum gap, indicating that its danger ranking transferred to the reactive follower.The measure predicted contact in 16.5% of forks versus 30.2% actual contact at the central follower braking rate.
- Calibration against a reactive follower: CCAR underestimated contact because the IDM follower braked more gently than the assumed b_F = 6 m/s².The calibration finding suggests that using a gentler assumed follower braking rate would raise predicted contact toward the realized rate.
8. Discussion
CCAR is an anticipatory exposure score for the blind region of proximity measures, but it is conditional rather than a crash-probability predictor. Its practical relevance is strongest for ego-follower forward-collision warning, while parameter assumptions, simulation sampling, and absent crash ground truth constrain interpretation.
- 8.1. Interpretation: CCAR scores exposure rather than crash probability, and trajectory data without crashes cannot validate it against observed outcomes.Its reported evidence combines correlation in observational data with controlled simulation evidence under leader braking.
- 8.2. Applicability to advanced driver-assistance systems: CCAR is more directly deployable for ego-follower forward-collision warning than for third-person research data.The host vehicle directly measures its own acceleration and senses leader gap and speed using mature radar-based systems.
- 8.2. Applicability to advanced driver-assistance systems: Operational deployment still requires an in-vehicle study of sensing latency, leader-state noise, and real-time computation.The paper makes no operational claim before those factors are addressed.
- 8.3. Limitations: CCAR values are conditional on assumed leader braking, reaction time, and follower braking parameters, so individual scores must travel with those assumptions.Only findings surviving the parameter sweep are reported.
- 8.3. Limitations: Spatial ranking by mean risk is parameter-sensitive, and simulation results depend on the sampled driver-parameter distribution rather than the dataset’s calibrated distribution.The paper treats only closing-acceleration-driven concentration as reliable across the reported analysis.
- 8.3. Limitations: The dataset’s absence of crashes provides no ground truth, motivating future evaluation on crash-related datasets.
9. Conclusions
The paper introduces and evaluates CCAR as a counterfactual closing-acceleration risk measure for longitudinal car-following. Across expressway trajectory data and controlled simulation, it represents states left undefined by proximity measures and links closing acceleration with collision risk under leader braking.
- Conclusions: 745,540 expressway trajectory frames were used to evaluate CCAR with controlled-simulation validation.
- Conclusions: CCAR assigns graded risk across the 51% of following states that proximity measures leave undefined.
- Conclusions: CCAR is not redundant with TTC, DRAC, or MTTC and responds to closing acceleration at fixed gap and speed.
- Conclusions: Controlled simulation confirms that closing acceleration is causally associated with collisions under leader braking.
APPENDIX
The appendix specifies the CCAR computation pipeline, baseline measures, evaluation tests, spatial analysis, and parameter robustness sweep. It also reports discriminator sensitivity across parameter settings.
- Parameter sensitivity: 66.33% is the maximum-discriminator result at b_L=8 m/s², t_r=1.5 s, and b_F=4 m/s², compared with 0.32% at the minimum-discriminator setting.The minimum setting is b_L=4 m/s², t_r=0.5 s, and b_F=8 m/s².
- Algorithm A1. CCAR computation pipeline: The pipeline audits trajectory quality, pairs leaders and followers by frame, excludes invalid observations, and computes per-frame baseline measures before CCAR.Inputs include vehicle identifiers, timing, lane, position, speed, acceleration, leader identifiers, and gap distance.
- Algorithm A1. CCAR computation pipeline: Baseline measures include TTC and DRAC only when the follower is closing, while MTTC uses the smallest positive root of a relative-motion quadratic with both vehicles’ accelerations.TTC is otherwise infinite and DRAC is otherwise zero.
- Evaluation design: The evaluation tests discrimination, rank correlation, top-decile risk-set overlap, within-cell acceleration trends, lane-level patterns, and parameter robustness.Robustness repeats the CCAR and analysis steps across combinations of leader braking, reaction time, and follower braking parameters.