Latency-Budgeted Headway: Sizing a Car's Following Gap by How Stale Its View of Hidden Road Users Is and How Distracted Its Driver Is

A car's safe following gap, and its safe speed past a parked truck, depends on one quantity that current systems mostly treat as a constant: how long it takes between a hazard appearing and the brakes engaging. Textbook stopping-distance formulas add a fixed reaction time to a braking term that grows with the square of speed. In practice that reaction time is set by at least two things that change from second to second. The first is how old the car's information is about road users hidden behind other vehicles, which it can only learn about from other cars or roadside sensors over a delayed, lossy radio link. The second is whether the human who must take over is actually taking in the road. We propose to measure both online and to feed them into a single latency budget that sets the car's headway and speed ceiling.

The core idea fits in two sentences. First, fuse the car's own fast sensing with slow, delayed shared detections using a complementary filter, which is a simple filter that trusts a fast sensor for quick changes and a slow sensor for long-run truth, and read the filter's built-in lag as the "information age" of each hidden region. Second, add a driver reaction term estimated from eye-movement scanning patterns rather than raw gaze direction, and size the safety envelope from the sum. The payoff we are testing is concrete: fewer conflicts with pedestrians who step out of occlusion shadows than latency-blind envelopes produce, and shorter gaps than a worst-case constant would force.

Core Claim

The literature has built the pieces separately. Velocity-scaled envelopes are well developed. PORA scales a stopping-sight-distance "safety box" with speed and weights risk by how fast occupancy probability changes. It reaches a conflict rate of 0.129 in mixed-traffic simulation, against 0.314 for time-to-collision. Its occupancy predictor, however, only covers what the car can see itself. Zhuyi derives how often each camera must be processed from stopping distance and tolerable latency. Its authors name occluded actors as the open problem, and the method assumes no human in the loop. The control-barrier-function work of Zeng et al. adds relative-velocity threat to the barrier and assumes full visibility.

Cooperative perception, in which vehicles and roadside units share sensor data over radio, does recover hidden actors. V2VNet improves occluded-object detection by over 60%. DAIR-V2X shows late fusion lifting LiDAR detection from 31.33 to 41.90 AP. Its time-compensated late fusion recovers only up to 1.5% AP under delay. All of these systems stop at detection output, and none of them says how old a shared detection may be before it can no longer be trusted for braking.

Driver-state work sits on the other side of the gap. Gao et al. find that fixation time barely tracks situational awareness (r = .12). Qu et al. predict awareness from scanning-graph features at 83.0% accuracy, though in a medical AR task. Visionary Co-Driver gates alerts on raw gaze, with about 2.29 s end-to-end latency. None of this work drives a control loop.

Our claim is that these belong in one inequality. For an occlusion shadow whose near edge is distance $d_s$ ahead, the speed $v$ is admissible only if

$$v\,\big(\tau_{\text{sys}} + \tau_{\text{info}}(t) + \tau_{\text{drv}}(t)\big) + \frac{v^2}{2a_{\max}} \le d_s + \Delta_{\text{coop}}(t),$$

where $\tau_{\text{sys}}$ is fixed actuation and processing delay and $a_{\max}$ is guaranteed deceleration. The term $\tau_{\text{info}}(t)$ is the effective information age of the shadow, taken from the fusion filter. The term $\tau_{\text{drv}}(t)$ is the driver's estimated reaction delay; it applies only in handover-dependent regimes, and in fully automated mode it is zero. The term $\Delta_{\text{coop}}(t)$ is the extra clear distance that fresh cooperative evidence certifies inside the shadow. That credit shrinks as evidence ages, at a rate set by the fastest plausible entry speed of a hidden pedestrian or cyclist.

The same latency sum replaces the constant reaction time in the ordinary following-gap rule for visible lead vehicles. What is new is not a velocity-scaled envelope. It is that two measured, time-varying delays, one from the network and one from the human, enter the envelope with physical units and no hand-tuned risk weights.

Research Program

The fusion layer runs one complementary filter per shadow cell of a bird's-eye-view grid. The car's own LiDAR recomputes the shadow geometry every frame, so it is the fast branch and passes through a high-pass filter: a cell that suddenly becomes visible updates immediately. Cooperative occupancy from other vehicles or a roadside unit is the slow branch and passes through the complementary low-pass filter.

Following Li's high-pass-first design, which shows that synthesizing the high-pass filter first and taking the low-pass as its exact complement gives much sharper rejection, we design the high-pass branch first. Its crossover frequency $\omega_c$ is adapted per cell from the measured message age and packet-loss rate. The static cutoff is a named limitation of standard complementary filters, and Sun et al. show that modeling the operating regime inside such a filter is what keeps it stable under extreme conditions.

For a first-order low-pass filter the group delay at low frequency is about $1/\omega_c$. This gives $\tau_{\text{info}} \approx \text{message age} + 1/\omega_c$, a closed-form, per-cell lag that costs a handful of operations. That matters because the envelope must be recomputed at control rate, and the rotation-vector orientation filter shows complementary-style updates running at 6.40 µs against 83.9 µs for a multiplicative extended Kalman filter.

The driver layer estimates $\tau_{\text{drv}}$. It maps a scanning-based awareness score, built on features in the spirit of Qu et al.'s fixation graphs (saccade amplitude and velocity, fixation dispersion, dwell on in-vehicle displays), to a reaction delay. That mapping is calibrated per participant on takeover trials in a driving simulator. Because Qu et al.'s 83.0% comes from a medical AR task, we treat transfer to driving as an open question and retrain on our own simulator data.

The interface layer is deliberately minimal and driven by the envelope. An alert for an occluded actor appears only when the inequality's slack falls below a threshold. This is a response to Gao et al.'s finding that highlighting in dense scenes lowers awareness (p = .02). We do not otherwise redesign the display.

The control layer enforces the inequality as a speed ceiling and gap constraint on an adaptive-cruise-style longitudinal controller. The same constraint can also enter a barrier-function controller as its safe set.

Baselines

The naive reference is a standard constant-time-gap adaptive cruise controller that uses the car's own sensing and a fixed reaction time. It is the do-nothing case that current Level 2 systems approximate. It runs on the identical CARLA scenarios and sensor suite, with its time gap swept over a range so we compare full safety-versus-efficiency curves rather than one hand-picked setting.

The strongest existing envelope method is PORA. Its speed-scaled safety box and occupancy-change hazard weighting are the closest published attempt to size risk dynamically. Its reported closed-loop number is an average conflict rate of 0.129, against 0.314 for time-to-collision and 0.262 for the Density Planner, in SUMO mixed traffic. We run PORA twice under the same data and tuning budget. The first run uses its native ego-only occupancy. The second is a stronger version that feeds it the same cooperative occupancy our method receives, treated as instantaneous. The second variant isolates our key ingredient, because it has cooperative information without age accounting.

The strongest cooperative-perception baseline is attentive intermediate fusion from OPV2V. It is reported at 0.734 AP at IoU 0.7 on default CARLA towns, against 0.406 without fusion. We pair it with DAIR-V2X's time-compensated late fusion as the delay-aware alternative. Each feeds a conventional envelope with a fixed reaction time.

Beating these baselines means a lower conflict rate with emerging occluded actors at equal or shorter mean time gap and equal or higher mean speed. That is a strictly better Pareto curve over the sweep, not a single operating point.

Three ablations isolate each ingredient. The first removes $\tau_{\text{info}}$ and treats cooperative data as fresh. The second fixes $\omega_c$ instead of adapting it. The third replaces the scanning-based $\tau_{\text{drv}}$ with an eyes-off-road timer. All systems get the same scenario seeds, channel traces and hyperparameter search budget.

Why not just…?

Why not just assume the worst-case delay and a permanently distracted driver? This is always safe and needs no estimation, so it is the real competitor on safety. Its cost is efficiency. Worst-case radio delay and worst-case human reaction, combined with the quadratic braking term, force low speed ceilings past every parked vehicle, which is the urban paralysis failure of static margins. Our Pareto sweep settles it directly. If the worst-case constant matches our conflict rate at a time gap within a few percent of ours, then online estimation is not worth it and we will report that.

Why not just use a Kalman filter with delayed-measurement handling instead of a complementary filter? A Kalman filter that handles out-of-order measurements may fuse at least as accurately, and this is a genuine threat to the filter choice. It is not a threat to the claim itself: the latency-budget inequality needs only some per-cell information age, and a Kalman filter's covariance could supply one. We run the Kalman variant as a drop-in. If it wins on conflicts, we adopt it and keep the complementary filter only where compute per cell dominates, which we measure across grid sizes.

Why not just gate on raw gaze or an eyes-off-road timer, as commercial driver monitors do? Gao et al.'s r = .12 between fixation and awareness says raw gaze underestimates reaction delay precisely for drivers who are looking but not seeing. That is the case where suppressing margin is most dangerous. Our third ablation and the human study compare the two estimators on how well they predict measured takeover time. If the timer predicts reaction time as well as the scanning score, we use the timer.

Why not just let PORA or a learned policy absorb the delays implicitly? A learned risk model trained on ego-only data has never seen stale cooperative evidence. The instantaneous-cooperative PORA variant tests whether implicit handling suffices. We expect it to fail specifically when message age exceeds a pedestrian's time to cross from the shadow edge into the lane.

Proposed Experiments

The automated study runs closed-loop in CARLA using OPV2V towns and the Culver City replica. We script occluded-pedestrian and cyclist emergence behind parked trucks and at blind intersections. Radio links are replayed from delay and packet-loss traces spanning from ideal to congested, with DAIR-V2X frames providing realistic cooperative detection quality for offline checks of the fusion layer. We found no reported closed-loop headway result on Lacuna for occlusion-emergence scenarios under cooperative delay, so all such numbers will be new.

Metrics are the conflict rate with emerging actors, minimum stopping margin, mean time gap and speed, and fusion-plus-envelope compute per control cycle. As an offline sanity check, the fusion layer's shadow-cell occupancy accuracy is compared against DAIR-V2X late fusion at the same delay.

The decisive check has two parts. We plot conflict rate against message age for all methods. Latency-blind methods should degrade sharply once age times pedestrian entry speed exceeds the lateral distance from shadow edge to lane, while ours should hold flat by slowing down. At the same time, our time-gap curve should stay well under the worst-case constant.

The human study uses a driving simulator in a Level 2 supervision regime, with a target of 24 participants. The protocol mirrors Gao et al., with Tobii-class eye tracking and SAGAT-style freeze probes, plus secondary-task distraction blocks. We measure takeover reaction time, awareness scores, alert count, and near misses. The decisive check is whether the scanning-based $\tau_{\text{drv}}$ predicts per-trial reaction time better than the eyes-off-road timer. It must also do so while the envelope issues fewer alerts than an always-highlight display in dense traffic.

If either decisive check fails, the corresponding term is dropped from the budget. The inequality is designed so that each term earns its place separately.