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

arXiv:2610.11783 · cs.LG · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "In-Ride Alcohol-Impairment Detection in E-Scooterists with False-Alarm Control".

Tom: The gist: Onboard sensors evaluate riders as they ride, raising an alarm as soon as enough evidence of impairment has accumulated,

Jane: First, who's behind it and why it matters.

Title and authors: Jane: We've got the full title now: "In-Ride Alcohol-Impairment Detection in E-Scooterists with False-Alarm Control". It really captures the essence of what they're doing, focusing not just on detection but also on controlling those false alarms.

Tom: I think that control aspect is key because if you get too many false alarms, people stop trusting the system altogether, which defeats the whole purpose >

Lu: The authors are Marco Capuccini and Rahul Rajendra Pai, and their work builds on previous research in automotive domains but applies those ideas to micromobility >

Meng: It’s interesting that they focused on fitting a logistic regression model for scoring at a lower computational cost compared to more complex alternatives, which is a practical engineering choice >

Lalam: They used the centered feature vectors of their training rides and took the impairment probability as the score, which sounds like a smart way to simplify things without losing too much accuracy >

Jane: So they’re essentially using a simpler modeling choice that works well enough for this application, while still being computationally efficient enough for deployment on hardware >

Tom: They show that this approach matches more complex alternatives at a lower computational cost, which is important when you're trying to deploy something on consumer-grade hardware like an e-scooter >

Lu: It highlights a trade-off between model complexity and the ability to run it in real time on limited resources, which is always a hot topic in AI >

Meng: From an engineering standpoint, that computational cost difference really matters for battery life and processing power on the device itself >

Lalam: It’s about making sure you have a system that’s both accurate enough to be useful and light enough to run smoothly on the hardware, which is a tough balance >

Jane: So it's not just about getting the right answer; it's about getting that right answer in a way that doesn't drain the battery or slow down your ride >

Tom: And this sets up perfectly for what we saw next: how this detection works continuously as you move through the ride >

The paper's summary: Tom: So, to get into the specifics of "In-Ride Alcohol-Impairment Detection in E-Scooterists with False-Alarm Control," the core contribution is introducing a detector that monitors e-scooter riding continuously and raises an alarm as soon as sufficient evidence accumulates >

Jane: Instead of waiting for a big event or a pre-ride test, the system is designed to react dynamically to the rider's actions during the trip >

Lu: This moves the focus from static checks to dynamic assessment, which is a really powerful shift in how we think about safety monitoring systems >

Meng: It means they are tracking kinematic data like accelerometer readings, gyroscope data, and throttle lever position measurements simultaneously over a short lookback window >

Lalam: They feed that seven-dimensional vector of these measurements into a machine learning model that spits out an impairment score for the current time window >

Tom: And then as the ride goes on, they use a running statistic of those scores to accumulate evidence, and an alarm triggers once that accumulated evidence crosses a threshold calibrated on sober rides >

Jane: This entire process is designed to remove the need for pre-ride screening and nighttime bans by only intervening when impairment is actually detected during the ride >

Lu: The key piece they introduced here is their threshold calibration procedure, which statistically bounds the false alarm rate at any step of the ride, and they prove its validity >

Tom: That statistical bounding means that even as you move along the ride, you know exactly how likely it is to get a false alarm at that moment >

Jane: And they show empirically that this detector keeps the false alarm rate within that bound where baselines that don't look at temporal dependence fail >

Lu: This is a big contribution because it shows why temporal dependence in the sensor data is so important for accurate impairment detection in this context >

Meng: So, to put it simply, they’re using continuous data and smart statistics to decide when to sound the alarm based on what's happening right now >

Lalam: It’s about making sure that every piece of evidence you collect is accounted for in the decision process, which builds a much stronger case than just looking at one snapshot >

The paper's improvements: Tom: Now we’re getting into what makes this approach an improvement over older methods. They aren't just proposing a new detector; they are showing how their method improves on existing ones.

Jane: The main improvements they highlight are that they show that their detector performs better in-ride under controlled false alarm control compared to baselines that don't consider temporal dependence >

Lu: Specifically, they show it outperforms these in-ride baselines when you keep the false alarm rate controlled while maintaining strong detection performance at the higher BAC level and moderate performance at the lower one >

Meng: That means this system is effective across a range of impairment levels, not just one extreme case, which is much more useful for real-world deployment >

Tom: And they proved that this detector keeps the false alarm rate within their bound throughout the ride—that's a guarantee we talked about earlier >

Jane: That guarantee means it’s not just good in an average sense; it's safe at every step of the ride because they prove its validity through their calibration procedure >

Lu: They show that this anytime valid bound follows directly, meaning the bound holds at every step of the ride rather than just at a single decision point >

Meng: So what does that mean for us? It means we don't have to worry about the system suddenly becoming unreliable halfway through a trip because it can be dependable continuously >

Tom: And finally, they also showed that this detector runs in real time on an embedded device, which is a big step toward making this actually usable in the field >

Conclusion: Jane: So to wrap up with "In-Ride Alcohol-Impairment Detection in E-Scooterists with False-Alarm Control," the main message is that you get a system that monitors riding continuously and raises an alarm based on accumulating evidence.

Tom: The real win here is the provable bound on the false alarms, which makes it reliable at every step of the ride, not just one pre-ride moment >

Lu: It means we can rely on this continuous monitoring to be dependable because they proved that their statistical method keeps things within bounds >

Meng: From a practical side, the real advantage is keeping the service available at night without having to implement those disruptive pre-ride screening processes or nighttime bans >

Lalam: It’s about giving riders a better experience by only intervening when impairment is actually detected, which sounds like a cleaner way to handle safety issues overall >

Jane: This paper on "In-Ride Alcohol-Impairment Detection in E-Scooterists with False-Alarm Control" shows that continuous monitoring combined with statistical control can provide a reliable tool for this problem >

Tom: It's a solid piece of work that moves us closer to having safety features built directly into the vehicle itself, and we’re looking forward to seeing what comes next >

Marco Capuccini, Rahul Rajendra Pai

KTH Royal Institute of Technology · Chalmers University of Technology

cs.LG

Submitted: 2026-10-08

Updated: 2026-10-08

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 82/100

The gist: The gist: Onboard sensors evaluate riders as they ride, raising an alarm as soon as enough evidence of impairment has accumulated, which removes the need for pre-ride tests and nighttime bans.

Key concepts

Detection Methodology: The system uses sensor readings (accelerometer, gyroscope, throttle) over short time windows to extract features. These features are fed into a machine learning model that assigns an impairment score for that specific window of riding.
Running Statistic: Evidence of impairment is accumulated by maintaining a running statistic of the scores generated from each ride window. An alarm is triggered immediately when this accumulated evidence crosses a pre-set threshold calibrated using data from sober rides.
Threshold Calibration: The system statistically sets the alarm threshold based on empirical data from sober rides. This calibration method ensures that the False Alarm Rate (FAR) is bounded at every step of the ride, providing a guarantee that alarms are reliable.
Real-Time Feasibility: The detector was tested on an embedded system, showing it can operate onboard an e-scooter. The detection cycle takes about 262 milliseconds, proving that the monitoring is fast enough for practical use while keeping data on the vehicle.

Summary

The gist: Onboard sensors evaluate riders as they ride, raising an alarm as soon as enough evidence of impairment has accumulated, which removes the need for pre-ride tests and nighttime bans.

Introduction and Motivation

Shared e-scooter services have become a widely adopted urban transport mode, yet alcohol intoxication stands out among the factors contributing to severe crashes Shared e-scooter services have become a widely adopted urban transport mode, yet alcohol intoxication stands out among the factors contributing to severe crashes Countermeasures remain limited to single-point reaction tests and night bans that suspend the service altogether In this paper, we propose an approach in which onboard sensors monitor rider kinematics continuously throughout the ride This approach removes the need for a pre-ride reaction test, keeps the service available at night, and enables in-ride intervention which occurs only when impairment is detected.

Detection Methodology

The detector operates on rides consisting of sensor readings, formally defined as a discrete-time signal R = 1 is a discrete-time signal of Algorithm 1 The detector evaluates the ride by extracting the PE of triaxial accelerometer, gyroscope, and throttle lever position measurements, computed over a short lookback window. The resulting seven-dimensional vector is then fed to a machine learning (ML) model that yields an impairment score for the current window. As the ride progresses, evidence is accumulated through a running statistic of the scores, and an alarm is raised as soon as it crosses a threshold calibrated on sober rides.

Threshold Calibration and Guarantee

The detector proposes a threshold calibration procedure that statistically bounds the FAR at any step of the ride, and we prove its validity The threshold is set to ta = S(⌈(K+1)(1−α)⌉), namely the empirical (1−α) quantile of the calibration peaks, where the count includes the ride yet to be observed and the rank is rounded up. The anytime-valid bound follows directly: P∃ j: Ej > ta = P S > ta.

Performance Evaluation

The evaluation relies on the dataset released with our prior work, comprising signals from 141 instrumented e-scooter rides, in which 25 participants rode while sober and at two target blood alcohol concentration (BAC) levels. The proposed detector keeps the FAR within the bound throughout, with point estimates consistently below α, and never significantly above it—as the lower CI ends show. Detection time is computed on each impaired ride raising an alarm, as w + (j∗ − 1)∆Fs, j∗ = min[j: Ej > ta], which amounts to the ride time elapsed when the alarming window j∗ closes.

Real-Time Feasibility

We evaluate an embedded implementation of the detector on an STM Nucleo board [10], where one detection cycle takes 262 ms, showing that in-ride monitoring is feasible onboard. This latency averages 262.26 ms, with a standard deviation of 0.91 ms and a range from 260 to 265 ms.

Conclusion

The paper presents an alcohol impairment detector that monitors e-scooterists as riding unfolds, raising an alarm as soon as enough evidence accumulates. The FAR is turned into a design parameter and bounded at every step of the ride, a guarantee we prove and confirm empirically. Detection is strong at the higher BAC level and moderate at the lower one, with alarms typically raised within half a minute of riding. In addition, we show that an embedded implementation leaves ample margin for realtime operation, allowing data to stay on the vehicle.

DATA AND CODE AVAILABILITY

All data and experimental code required to reproduce the reported results are publicly available at https://github.com/voi-oss/in-ride-impairment-detection and https://doi.org/10.6084/m9.figshare.34070439 The authors have no competing interests to declare that are relevant to the content of this article. The data used in this paper come from the experiment reported in our prior work [8], run at MicroLab, Chalmers University of Technology, Sweden. We sincerely thank the participants for their time, Marco Dozza, Niyathi Kini, Alexander Rasch, Ali Mohammadi from Chalmers for the data collection. The data collection experiment was supported by the MicroTox project, funded by VINNOVA (Sweden’s innovation agency) and Drive Sweden under grant 2025-00431. This work was partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. The authors take full responsibility for the content of this paper. The use of generative AI tools was used to assist with information retrieval, experimental code development, and text editing.

REFERENCES

[1] Fluctuo, “European shared mobility annual review 2025,” 2026, accessed: Jul. 31, 2026 [Online]. Available: https://fluctuo.com/pdf/e smi-2025.pdf [1] H Stigson, I Malakuti, and M Klingegard, “Electric scooters accidents: Analyses of two Swedish accident data sets,” Accident Analysis & Prevention, vol. 163, p. 106466, Dec. 2021 [Online]. Available: http://dx.doi.org/10.1016/j.aap.2021.106466 [3] R R Pai, R Fredriksson, and M Dozza, “Three modes, three profiles: Characterizing fatal crashes on e-scooters, e-bikes, and conventional bicycles in Sweden,” Journal of Safety Research, vol. 97, p. 533–542, Jun. 2026 [Online]. Available: http://dx.doi.org/10.1016/j.jsr.2026.05. [4] J B Cicchino, P E Kulie, and M L McCarthy, “Severity of e-scooter rider injuries associated with trip characteristics,” Journal of Safety Research, vol. 76, p. 256–261, Feb. 2021 [Online]. Available: http://dx.doi.org/10.1016/j.jsr. [5] F Hartz, P Zehnder, T Resch, G Rommermann, V Hartmann, M Schwarz, C Kirchhoff, P Biberthaler, and M Zyskowski, “Characteristics of e-scooter and bicycle injuries at a university hospital in a large German city – a one-year analysis,” Injury Epidemiology, vol. 12, no. 1, Jan. 2025 [Online]. Available: http://dx.doi.org/10.1186/s40621-024-00554-w [6] S Dibaj, S Vosough, K Kazemzadeh, S O’Hern, and M N Mladenovic, “An exploration of e-scooter injuries and severity: Impact of restriction policies in Helsinki, Finland,” Journal of Safety Research, vol. 91, p. 271–282, Dec. 2024 [Online]. Available: http://dx.doi.org/10.1016/j. [7] Voi Technology, “Tap the helmets: Voi introduces world’s first Reaction Test for e-scooters to discourage drunk riding,” Voi Technology blog, Sep. 2020, accessed: Jul.

Improvements for AI systems

Bold header: Anytime-Valid False-Alarm Control

The system can raise an alarm as soon as enough evidence accumulates without inflating false alarms by using a threshold calibrated on sober rides to bound the rate of incorrect detections at every step of the ride, as proven by the claim: Anytime validity means that the bound holds at every step of the ride rather than at a single pre-specified decision point.

Bold header: Continuous, Onboard Monitoring

The AI system can monitor rider kinematics continuously throughout the trip to raise an alarm as soon as enough evidence of impairment has accumulated, which removes the need for a pre-ride reaction test and enables intervention onboard—without requiring data to leave the vehicle.

Bold header: Real-Time Embedded Implementation

The system is feasible for deployment in physical devices because it meets the realtime requirement, with an embedded implementation on an STM Nucleo board taking only 262 ms per detection cycle, ensuring that in-ride monitoring is feasible onboard.

Sources

Kinematic signatures of impairment: Detecting alcohol intoxication in e-scooter riders using sensor data and machine learning