PaperlayerView exact revision on arXiv

Source-linked AI summary

In-Ride Alcohol-Impairment Detection in E-Scooterists with False-Alarm Control

Marco Capuccini, Rahul Rajendra Pai

arXiv:2610.11783v1cs.LG

TL;DR

Alcohol impairment contributes to e-scooter crashes, but existing countermeasures rely on start-of-trip tests or nighttime bans. This paper introduces continuous onboard sensing with an anytime-valid false-alarm bound; at a bound of 0.023, it detects 91% of higher-BAC rides and 50% of lower-BAC rides.

  • Problem

    E-scooter impairment countermeasures remain limited to easily circumvented start-of-trip reaction tests and nighttime bans that suspend service for everyone.

    Source: §I · p. 1

  • Method

    The detector continuously scores windowed inertial and throttle measurements, accumulates evidence, and uses a calibrated threshold with a provable FAR bound at every ride step.

    Source: §I · pp. 1–2

  • Results

    At a FAR bound of 0.023, it detects 91% of higher-BAC rides within a median 25 s and 50% of lower-BAC rides within a median 27 s.

    Source: §I · p. 1

  • Takeaways & Limitations

    Onboard anytime-valid alarms offer an alternative to pre-ride screening and nighttime bans while letting operators budget false alarms for sober riders.

    Source: §I · p. 1

    Detection at BAC levels of 0.12%–0.20% remains untested because inducing that degree of intoxication could expose participants to higher injury risk.

    Source: §VI · p. 6

Abstract

from arXiv · show

Shared e-scooter services have become a widely adopted urban transport mode. While most users ride responsibly, alcohol intoxication stands out among the factors contributing to severe crashes. Nonetheless, countermeasures remain limited to single-point reaction tests and night bans that suspend the service altogether. This paper proposes a new approach in which onboard sensors evaluate the rider as the trip unfolds, raising an alarm as soon as enough evidence of impairment has accumulated. Specifically, we introduce a detector that operates on inertial and throttle measurements, with a provable bound on the rate of false alarms. Experiments on sensor data from 141 rides, in which 25 participants rode while sober and at two target blood alcohol concentration levels, confirm that the bound holds, whereas baselines and ablations either exceed it or lose detection performance, and in some cases delay the alarm. At a bound of 0.023, the detector identifies 91% of the rides performed at the higher concentration and 50% of those at the lower one, with median detection times of 25 and 27 seconds, respectively. We further show that an embedded implementation meets the real-time requirement, making mitigation actions feasible onboard, without requiring data to leave the vehicle. Overall, this work lays the ground for interventions that reach impaired riders as soon as possible, sparing the sober ones the burden of a pre-ride test or the suspension of the service at night, while letting operators budget false alarms against user experience.

I. INTRODUCTION

Alcohol-related crashes make e-scooter safety a pressing concern, while existing countermeasures are limited to start-of-trip tests and nighttime bans. The paper proposes continuous onboard monitoring with a statistically controlled alarm and evaluates its detection performance.

I. INTRODUCTION

  • Existing reaction tests screen riders only at trip start and are easily circumvented, while nighttime bans suspend service for sober riders too.Source: §I · p. 1
  • Onboard sensors monitor rider kinematics throughout the ride, accumulating evidence until an impairment alarm is warranted.The detector extracts permutation entropy from inertial and throttle measurements and feeds the resulting features to a machine-learning model.Source: §I · p. 1
  • At a FAR bound of 0.023, the detector identifies 91% of higher-BAC rides within a median 25 s and 50% of lower-BAC rides within a median 27 s.No in-ride baseline matches its detection rate under FAR control.Source: §I · p. 1
  • The proposed calibration procedure bounds the false-alarm rate at every step of the ride, and the paper proves its validity.Anytime validity permits an alarm as soon as sufficient evidence accumulates, without inflating the FAR.Source: §I · pp. 1–2

II. RELATED WORK

Prior work studies alcohol impairment in vehicles and micromobility, but anytime-valid false-alarm control for in-trip detection remains unexplored. The detector uses filtered, windowed scooter signals and rider-centered features.

II. RELATED WORK

  • In-ride anytime-valid FAR control remains unexplored across automotive and micromobility domains, while temporal dependence can invalidate exchangeability-based guarantees for sensor streams.Earlier approaches either do not address dependence or incur detection delay through shuffling; other conformal approaches target single-point coverage rather than anytime validity.Source: §II · p. 2

III. DETECTION METHOD

  • The detector uses seven sensor channels—triaxial acceleration, triaxial angular velocity, and throttle position—filters inertial readings, and retains a 1,000-sample lookback window, updated every 100 samples.The inertial readings pass through a causal 64th-order finite impulse response low-pass filter using a Hamming window at a 10 Hz cutoff.Source: §III · p. 2
  • Each window is represented by normalized permutation entropy features, then centered using channel-wise means from rides by the same rider.Centering removes a rider-specific offset without requiring knowledge of impairment level.Source: §III · p. 2

A. Inference

Inference scores each ride window and accumulates the scores as the ride unfolds. An alarm is raised when the running evidence exceeds a calibrated threshold.

A. Inference

  • Each window receives an impairment score from logistic regression fitted on centered feature vectors from training rides with known impairment labels.Higher scores indicate stronger evidence of impairment; training rides may come from riders other than the evaluated rider.Source: A. Inference · p. 3
  • The running evidence is a shrunk average of window scores, with κ = 10 pseudo-windows pulling early estimates toward the sober baseline.This shrinkage damps small-sample fluctuations early in the ride.Source: A. Inference · p. 3
  • The procedure extracts and centers features, scores each window, and updates running evidence; it alarms above the threshold or continues to the next window.Because the shrunk average can be computed from a running sum, resource use does not grow with the number of windows.Source: A. Inference · p. 3

B. Threshold Calibration

Threshold calibration processes separate sober rides and sets the alarm threshold from their peak running evidence. The resulting test bounds the FAR at every ride step.

B. Threshold Calibration

  • Calibration rides are sober, randomly sampled, and disjoint from training and centering rides; each is processed without impairment testing.Source: B. Threshold Calibration · p. 3
  • The threshold is the empirical (1−α) quantile of calibration-ride peaks, with the ride under test included in the count and the rank rounded up.Source: B. Threshold Calibration · p. 3
  • Testing against the calibrated threshold keeps the FAR within α at every step of the ride.Source: B. Threshold Calibration · p. 3

C. Anytime-Valid False-Alarm Guarantee

The guarantee extends false-alarm control from a single test point to every step of a ride by calibrating against each ride’s peak running evidence. Its validity relies on exchangeability of the calibration rides and the ride under test.

C. Anytime-Valid False-Alarm Guarantee

  • The ride-level peak converts a single-point ICP validity proof into a bound covering every step of the ride.A threshold crossing at any step is equivalent to the peak exceeding the threshold.Source: C. Anytime-Valid False-Alarm Guarantee · p. 3
  • Exchangeability makes the peak equally likely to occupy any of the K + 1 ranks, supporting the calibration bound.The rides’ peaks inherit exchangeability because they are deterministic functions of the rides, with random sampling for calibration rides and an assumption for the ride under test.Source: C. Anytime-Valid False-Alarm Guarantee · p. 3
  • The authors state that the anytime-valid bound follows directly from the rank argument.Source: C. Anytime-Valid False-Alarm Guarantee · p. 3

IV. DETECTION PERFORMANCE

This section empirically evaluates detection performance after the preceding section establishes a FAR bound by construction. It examines whether impairment is detected and how quickly, comparing baselines with and without FAR-control guarantees.

IV. DETECTION PERFORMANCE

  • The evaluation tests detection rate and detection time because the FAR guarantee alone does not establish how often or how quickly impairment is detected.Source: §IV · p. 3
  • The section evaluates detection performance against baselines both with and without FAR-control guarantees.Source: §IV · p. 3
  • The evaluation proceeds from dataset, metrics, experimental setup, and baselines to the results.Source: §IV · p. 3

A. Dataset

The evaluation uses 141 rides from 25 participants in a controlled indoor e-scooter experiment, with sober and two target BAC conditions.

A. Dataset

  • The final dataset contains 141 rides: 45 sober, 50 at 0.05% BAC, and 46 at 0.08% BAC, from 25 participants.At least one ride per participant remained in each condition.Source: A. Dataset · p. 3
  • Participants rode a 7 km/h-limited scooter on an indoor track and completed two sober trials followed by two trials at each BAC level.The target BAC levels were 0.05% and 0.08%, confirmed by breathalyzer before every trial.Source: A. Dataset · p. 3
  • Logger memory exhaustion excluded eight participants and caused nine lost trials across eight others, leaving 25 participants in the analysis.Recordings had fewer than 10% missing samples, filled by linear interpolation.Source: A. Dataset · p. 3

B. Evaluation Metrics

Performance is measured through false-alarm rate, detection rate, and detection time, with detection time assessed on impaired rides that raise an alarm. The setup uses leave-one-participant-out scoring and calibration, with thresholds evaluated at feasible FAR bounds.

B. Evaluation Metrics

  • FAR and detection rate are alarm rates on sober and impaired rides, respectively; detection time is measured only on impaired rides that alarm.Source: B. Evaluation Metrics · p. 4

C. Experimental Setup

  • In leave-one-participant-out evaluation, each ride is scored out of fold, with a participant-specific threshold calibrated from held-out sober-ride peaks.The scoring function is fit on the other participants’ rides, and their sober rides provide the baseline.Source: C. Experimental Setup · p. 4
  • FAR and detection-rate confidence intervals use 1000 participant-clustered bootstrap resamples, while detection time is summarized by its median and IQR.Source: C. Experimental Setup · p. 4
  • The pooled FAR bound relies on exchangeability of calibration peaks, while the authors argue rider-specific centering makes the held-out-participant distinction negligible.They extend the same reasoning to confidence intervals, which resample participants at random.Source: C. Experimental Setup · p. 4
  • The smallest attainable α is about 0.023 because holding out a participant leaves at least K = 43 sober calibration rides.The reported α values are 0.023, 0.05, 0.1, and 0.2.Source: C. Experimental Setup · p. 4

D. Baselines

The baselines isolate the effects of whole-ride versus in-ride scoring, running evidence, and threshold calibration, alongside a sequential martingale alternative.

D. Baselines

  • The comparison fixes the window size at 1,000 samples to balance PE estimation accuracy against minimum detection delay.The evaluation also includes ablations of in-ride detection and calibrated running evidence, then tests exchangeability martingales and shrinkage.Source: D. Baselines · p. 4

1) Batch Classifier:

  • The batch classifier scores whole-ride PE and flags rides above 1 −α without calibration, so its alarm cannot precede ride completion.This reproduces the prior work [8].Source: 1) Batch Classifier · p. 4

2) Naive Windowed Classifier:

  • The naive windowed classifier scores each window independently and flags impairment when any score exceeds 1 −α, without running evidence or calibration.Source: 2) Naive Windowed Classifier · p. 4

3) Exchangeability Martingale:

  • The exchangeability martingale sequentially bets on smoothed conformal p-values, testing its wealth process against 1/α under Ville’s inequality.Its jump probability is 0.01.Source: 3) Exchangeability Martingale · p. 4

4) Recalibrated Martingale:

  • The recalibrated martingale uses the same wealth process as running evidence but calibrates its threshold using the proposed procedure.Source: 4) Recalibrated Martingale · p. 4

5) Shrinkage Sweep:

The results compare detection and false-alarm control across baselines, then examine how shrinkage changes detection while preserving the observed FAR.

E. Results

  • The proposed detector stays within all four FAR bounds, while naive windowed and exchangeability-martingale baselines significantly inflate FAR.The proposed detector’s point estimates stay below α, though upper confidence limits can exceed it; the recalibrated martingale has a statistically insignificant 0.11 estimate at α = 0.1.Source: E. Results · pp. 4–5, Table I · p. 5
  • At α = 0.023, the detector reaches 91% at higher BAC (25 s median) and 50% at lower BAC (27 s), versus recalibrated-martingale rates of 65% and 14%.The comparison is for methods that control FAR; Table II also reports 95% CIs and IQRs.Source: Table II · p. 5, E. Results · p. 5

5) Shrinkage Sweep:

  • With κ = 0, detection falls to 63% and 22% despite 10 s alarms; rates peak at κ = 10, then fall to 85% and 38% at κ = 30, while FAR stays 0.02.The paired detection rates are for higher and lower BAC, respectively; κ = 0 removes the pull toward the sober baseline.Source: 5) Shrinkage Sweep · p. 4, E. Results · p. 5

V. ON-DEVICE LATENCY

The embedded implementation meets the timing margin for real-time operation, while the paper identifies constraints on its evaluation and on how alarms should be acted upon.

V. ON-DEVICE LATENCY

  • Over 1,000 cycles, latency averaged 262.26 ms (standard deviation 0.91 ms; range 260–265 ms), taking about a quarter of the second separating consecutive windows.The evaluation timed Algorithm 1 in C, bare metal on a NUCLEO-H533RE at the default 32 MHz system clock.Source: §V · p. 5

VI. DISCUSSION AND LIMITATIONS

  • Calibration uses sober outcomes, avoiding a need for impaired training rides; rider centering needs only a handful of prior unlabeled rides.The paper notes that impaired rides are scarce because of limited reporting and ethical concerns.Source: §VI · pp. 5–6
  • An alarm indicates an atypical pattern relative to the sober baseline, not intoxication, so the paper recommends soft measures rather than accusation or sanction.Interventions should be tiered, transparent, and subject to human review when costly; onboard measures need not send data off the vehicle.Source: §VI · p. 6
  • Evaluation uses a small dataset and fixed maneuvers on one vehicle; higher BAC detection is stronger, while more severe levels remain untested.Naturalistic labels are out of reach under controlled dosing, and traffic, slopes, and surfaces may raise sober peaks or mask impairment; per-vehicle and location recalibration is described as inexpensive.Source: §VI · p. 6

VII. CONCLUSION

  • The conclusion presents in-ride alarms as a foundation for interventions during trips, reducing the need for pre-ride screening and nighttime bans.The embedded implementation allows data to stay on the vehicle.Source: §VII · p. 6

Related papers

Loading 2610.11783v1…