⚡ electrical engineering

Counterfactual Closing-Acceleration Risk: An Anticipatory Surrogate Safety Measure for the Blind Region of Car-Following

This paper introduces Counterfactual Closing-Acceleration Risk (CCAR), an anticipatory surrogate safety measure that quantifies rear-end collision risk based on a follower's gap-closing acceleration to address the "blind region" where traditional metrics like time-to-collision remain undefined, thereby providing graded risk assessments for the majority of car-following scenarios previously left unscored.

Original authors: Eni Solomon Laughter

Published 2026-09-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Eni Solomon Laughter

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Road safety research has long faced a frustrating paradox: the most dangerous moments on the road often leave no trace. Because actual crashes are rare events compared to the billions of miles driven safely every day, scientists cannot simply wait for accidents to happen to understand how to prevent them. Instead, they rely on "surrogate safety measures," which are mathematical tools designed to spot near-misses. These tools look at how close two cars come to hitting each other without actually colliding, treating that near-miss as a warning sign. The most common of these tools, known as time-to-collision, works on a simple assumption: it calculates how long it would take for a rear-end crash to happen if both drivers kept their current speed and direction. This works well when a car is already speeding up toward the vehicle in front of it. However, this method has a blind spot. It cannot see danger when a driver is approaching a slower car but is still traveling at a slower speed themselves, even if they are pressing the gas pedal to close the gap. In those moments, the standard tools declare the situation safe because the gap is not shrinking fast enough yet, ignoring the fact that the driver is building up speed that could become dangerous in an instant.

A researcher at Chang'an University in Xi'an, China, has developed a new way to look at this hidden danger. They call their new measure Counterfactual Closing-Acceleration Risk. The core idea is to ask a "what if" question that standard tools never ask: what would happen if the car in front suddenly slammed on its brakes right now? The researcher realized that a driver who is currently accelerating to catch up to a leader is accumulating a specific kind of risk that has nothing to do with their current speed or distance, but everything to do with how hard they are pressing the gas. By projecting this hypothetical braking event forward, the new measure can score the danger of a situation even when the following car is not yet faster than the one ahead. This approach turns the "blind" moments of driving into a readable map of risk, revealing that nearly half of all car-following situations on highways are currently invisible to traditional safety tools.

To test this idea, the researcher analyzed a massive collection of real-world driving data from a busy expressway in Nanjing, China. This dataset, captured by aerial cameras, contained over 745,000 individual moments of cars following one another. When they applied their new method to this data, they found that standard tools left 51 percent of these moments un-scored, effectively treating them as safe because the following car was not yet overtaking the leader. In contrast, the new measure found that 98.8 percent of these "blind" moments returned a graded, non-trivial risk score, indicating that the follower's current acceleration created a measurable exposure to danger. Furthermore, 7.6 percent of these blind-region frames produced a counterfactual collision, meaning that in the mathematical projection, the minimum gap between the vehicles dropped to zero or below if the leader had braked. The researcher also checked how this new measure compared to existing ones and found they were not just repeating the same information. While the new measure agreed somewhat with older tools, it flagged different moments as dangerous, proving that the act of accelerating to close a gap provides unique warning signs that distance and speed alone cannot offer.

To be sure that this pattern was not just a statistical coincidence, the researcher built a controlled computer simulation. They created a virtual highway scene with a follower car programmed to drive realistically and a leader car that would brake suddenly at a specific moment. They ran thousands of these scenarios, changing only how hard the follower was accelerating when the leader braked. The results were clear: when the gap and the leader's braking force were held constant, the actual rate of collisions rose steadily as the follower's acceleration increased. This confirmed that the acceleration itself was a direct cause of the danger, not just a side effect. The simulation showed that a driver who is aggressively closing a gap is far more likely to crash if the car ahead stops suddenly than a driver who is simply coasting, even if both drivers are at the same distance and speed at the moment the brakes are hit.

The study also looked at where these risks were most likely to occur on the road. The data showed that the danger was not spread evenly but was concentrated in specific areas, particularly in the lanes where cars merge onto the highway or change lanes. These are the zones where drivers naturally accelerate to claim a gap, making them the most vulnerable spots for the type of risk the new measure detects. The researcher tested their method against different assumptions about how fast cars can brake and how long it takes a driver to react, finding that the results remained consistent across a wide range of realistic scenarios. While the new tool is not a perfect crystal ball that predicts every crash, it offers a continuous stream of risk information where previous tools saw nothing. It suggests that for advanced safety systems in cars, paying attention to how hard a driver is pressing the accelerator could be just as important as watching how far away the car in front is, potentially allowing vehicles to warn drivers or brake automatically before a dangerous situation becomes unavoidable.

Technical Summary: Counterfactual Closing-Acceleration Risk (CCAR)

Problem Statement
Current surrogate safety measures (SSMs) for road safety assessment, such as Time-to-Collision (TTC), Deceleration Rate to Avoid a Crash (DRAC), and Modified TTC (MTTC), rely heavily on proximity and constant-velocity assumptions. A critical limitation of these dominant measures is their "blindness" to latent risk when the following vehicle is not yet traveling faster than the leader. In these states, relative speed is non-positive, rendering TTC infinite and DRAC zero, thereby classifying the interaction as safe. However, this classification ignores the follower's acceleration state. If a follower is accelerating to close a gap, a sudden braking event by the leader could lead to a collision before the follower can react. Existing measures fail to distinguish between a follower maintaining a constant speed and one aggressively accelerating into a gap, despite the latter carrying significantly higher latent risk.

Methodology
The paper introduces Counterfactual Closing-Acceleration Risk (CCAR), an anticipatory surrogate measure designed to quantify exposure in these "blind" regions.

Conceptual Framework: CCAR answers a counterfactual question: If the leader were to brake immediately, how close would the vehicles come, given the follower's current gap-closing acceleration?
Mathematical Construction: The measure projects a hypothetical scenario starting at the evaluation instant.
- Leader: Assumes immediate constant deceleration ( $b_L$ ) until stop.
- Follower: Maintains current closing acceleration ( $\alpha$ ) through a reaction time window ( $t_r$ ), then applies maximum braking capability ( $b_F$ ).
- Risk Calculation: The minimum projected gap ( $s_{min}$ ) is calculated analytically by evaluating the gap function at stationary points and phase boundaries. The risk score ( $R$ ) is a normalized value between 0 and 1, derived from the ratio of the minimum gap to the initial gap. A counterfactual collision is defined as $s_{min} \leq 0$ .
Data and Validation:
- Trajectory Analysis: The measure was evaluated on 745,540 car-following frames from the SQM-W-1 dataset (Nanjing, China), a high-resolution aerial trajectory dataset sampled at 24 Hz.
- Parameter Sensitivity: The model treats leader braking rate ( $b_L$ ), reaction time ( $t_r$ ), and follower braking capability ( $b_F$ ) as swept sensitivity parameters (27 combinations) rather than fixed constants to ensure robustness.
- Simulation: A controlled simulation using the Intelligent Driver Model (IDM) with scripted leader braking was employed to establish causal links. This setup allowed for the observation of actual collision rates under varying closing accelerations while holding gap and leader braking constant.

Key Contributions

Formalization of CCAR: The study defines CCAR as an objective, anticipatory, counterfactual SSM specifically for longitudinal car-following.
Addressing the "Blind Region": The paper demonstrates that CCAR provides a graded, non-trivial risk score for approximately 51% of car-following states (where the follower is not faster than the leader) that remain unscored or deemed safe by TTC and DRAC.
Informational Value of Acceleration: It establishes that follower closing acceleration carries risk information distinct from gap and relative speed. CCAR is not redundant with existing measures (Spearman correlation with MTTC is 0.54; top-decile risk-set overlap is 0.19).
Causal Validation: Through controlled simulation, the paper provides evidence that, holding gap and leader braking fixed, the actual collision rate of a reactive follower rises monotonically with closing acceleration, supporting the causal role of the precursor.

Results

Prevalence of Latent Risk: In the "blind region" (51.2% of frames), CCAR returns a non-zero risk in 98.8% of cases, with a median risk score of 0.485. At a standard TTC threshold of 1.25s, 14.4% of frames classified as "safe" by TTC produced a counterfactual collision under CCAR.
Spatial Distribution: Counterfactual collision risks are not uniform; they concentrate in transition and partial lanes (lanes 3, 4, and 5) within merge-influence zones, aligning with documented gap-closing maneuvers.
Robustness: The discriminator set (states where TTC is safe but CCAR is risky) remains non-zero across all 27 parameter settings, ranging from 0.3% to 66.4% of frames. The correlation with MTTC remains moderate (0.36–0.62) across settings, confirming the measure's independence.
Simulation Outcomes: In the IDM simulation, the contact rate increased from 0.22 to 0.60 as closing acceleration bands increased, confirming the causal association between closing acceleration and collision risk under leader braking.

Significance and Claims
The paper positions CCAR not as a replacement for crash prediction, but as a continuous risk label that exists where conventional measures are silent. Its primary significance lies in:

Anticipatory Capability: It scores exposure based on the follower's current motion state (acceleration) rather than just the time-to-impact under current velocities.
Applicability to ADAS: The measure is particularly suited for Forward Collision Warning (FCW) systems in the ego-vehicle setting, where the host vehicle can directly measure its own speed and acceleration, while sensing the leader's gap and speed.
Building Block: CCAR is proposed as a reusable feature for behavior and prediction models and a screening tool for surfacing latent-risk states that would otherwise be ignored.

The author explicitly notes limitations: the measure is conditional on assumed parameters ( $b_L, t_r, b_F$ ), does not predict actual crashes (as trajectory data lacks crash outcomes), and is restricted to longitudinal dynamics. Future work is suggested to test the model on crash-related datasets and address real-time sensing latency.

Technical Summary: Counterfactual Closing-Acceleration Risk (CCAR)

More like this