Source-linked AI summary
Driver Identification Using Automobile Sensor Data from a Single Turn
David Hallac, Abhijit Sharang, Rainer Stahlmann, Andreas Lamprecht, Markus Huber, Martin Roehder, Rok Sosic, Jure Leskovec
TL;DR
The paper studies whether drivers can be identified from a short segment of vehicle sensor data, such as one turn. It applies a classifier to repeated turn recordings from Audi vehicles and reports accurate identification across 12 common turns, with results varying by turn and setting.
Problem
The paper asks whether a driver’s identity can be inferred from only a short road segment, where sensor readings may reveal individual driving patterns.
Source: §I · p. 1Method
The authors develop a classifier for driver identity from sensor time series collected repeatedly at a single turn and evaluate it on Audi real-road data.
Source: §I · p. 2, §III · p. 3Results
Average single-turn prediction accuracy was 76.9% for two-driver classification and 50.1% for five-driver classification across the 12 most frequent turns.
Source: §V · p. 5Takeaways & Limitations
Turns provide multiple driver actions within seconds, and the authors report that rural locations appear easier to classify than urban ones.
Source: §I · pp. 1–2The evaluation covers the 12 most frequently made turns in the Audi dataset; testing the method on longer time series is left for future work.
Source: §I · p. 2, §VI · p. 6
Abstract
from arXiv · showhide
As automotive electronics continue to advance, cars are becoming more and more reliant on sensors to perform everyday driving operations. These sensors are omnipresent and help the car navigate, reduce accidents, and provide comfortable rides. However, they can also be used to learn about the drivers themselves. In this paper, we propose a method to predict, from sensor data collected at a single turn, the identity of a driver out of a given set of individuals. We cast the problem in terms of time series classification, where our dataset contains sensor readings at one turn, repeated several times by multiple drivers. We build a classifier to find unique patterns in each individual's driving style, which are visible in the data even on such a short road segment. To test our approach, we analyze a new dataset collected by AUDI AG and Audi Electronics Venture, where a fleet of test vehicles was equipped with automotive data loggers storing all sensor readings on real roads. We show that turns are particularly well-suited for detecting variations across drivers, especially when compared to straightaways. We then focus on the 12 most frequently made turns in the dataset, which include rural, urban, highway on-ramps, and more, obtaining accurate identification results and learning useful insights about driver behavior in a variety of settings.
I. INTRODUCTION
The paper asks whether a driver can be identified from a few seconds of vehicle sensor data, where individual driving styles may leave distinctive sensor patterns. It develops a classifier for this task and evaluates it on Audi data, focusing on turns as short segments with multiple driver actions.
I. INTRODUCTION
- Short-segment identification could support early driver recognition, locally tailored vehicle settings, driver profiles, and road-segment warnings without sending data off-car.The authors describe these as potential applications and say local implementation keeps driver information private.Source: §I · p. 1
- Turns capture several driver actions within seconds, while straightaways contain fewer distinct actions for distinguishing drivers.The authors note that repeated drives may vary with pedestrians or driver mood, so they focus on segments where prediction accuracy is high.Source: §I · p. 1
- On the 12 most frequent turns, average prediction accuracy is 76.9% for two-driver classification and 50.1% for five-driver classification.The authors also report that rural locations appear easier to classify than urban ones, likely because they are more consistent across sessions and have fewer potential distractions.Source: §I · p. 2
- The method identifies a driver from a short segment such as a single turn, using patterns in vehicular sensor data.The paper’s contributions include a new real-road Audi dataset and an algorithm that predicts among a given set of individuals.Source: §I · p. 2
- Earlier automobile-sensor work modeled driving behavior and identified drivers, but the cited approaches addressed longer sessions or omitted time-series temporal and spectral aspects.The paper contrasts its few-second single-turn task with a prior shortest test of one minute and describes its classifier as using additional sensors, different features, and a larger, more diverse dataset.Source: §I · p. 2
II. DATASET DESCRIPTION
The dataset was collected from retrofitted Audi vehicles driven on real roads near Ingolstadt, and includes 64 drivers across 10 cars. The analysis selects and checks the 12 most frequent turns, with turns operationally defined by heading change, duration, and preceding stability.
II. DATASET DESCRIPTION
- Contractors drove retrofitted Audi A3 vehicles that logged sensor readings on real roads in and around Ingolstadt, Germany.Timestamped CAN-bus data was stored when individual sensor values changed, rather than at synchronous intervals.Source: §II · p. 2
- The dataset contains 10 cars, 64 drivers, 2,098 driving hours, 110,023 kilometers, and 1,889 driving sessions.Sessions were continuous intervals of active driving, typically three to four per day, though the number varied widely.Source: §II · p. 2
- A turn requires at least 70 degrees of heading change, duration under 10 seconds, and at least 5 seconds of stable heading beforehand.These conditions are intended to reduce false positives from curvy roads or parking lots.Source: §II · p. 2
- The 12 most common turns were manually checked and labeled by type, spanning city roads, rural areas, and highway ramps.The most frequent turn is the first turn outside the Audi parking lot, aligning with the goal of identifying drivers at the beginning of a drive.Source: §II · p. 3
- For one example turn, drivers differed in when they began turning, peak steering-wheel angle, and how long they held the wheel turned.The illustration compares average steering-wheel angle across sessions for three drivers, each of whom made that turn 12 to 18 times.Source: §II · p. 3
III. CLASSIFICATION ALGORITHM
The method classifies a driver from sensor readings collected during a short road segment. It aligns sessions by location, then represents each sensor with statistical and wavelet-derived features for a random-forest classifier.
III. CLASSIFICATION ALGORITHM
- Each session covers a 150-foot radius around a centerpoint, typically about 10 seconds, with 12 sensor readings sampled every 0.1 seconds.At turns, these time intervals are usually 1–3 feet apart.Source: §III · pp. 3–4
- Sessions are aligned by location rather than elapsed time, using a baseline session’s sampled GPS locations and weighted averages of neighboring readings.The baseline is chosen to minimize changes in velocity; the resulting aligned series contains approximately 100 locations per session.Source: §III · p. 4
- For each sensor, seven summary statistics and five PCA-reduced Haar-wavelet features form the 12-feature representation supplied to a random forest.The statistics include mean, standard deviation, skew, kurtosis, minimum, maximum, and autocorrelation.Source: §III · p. 4
IV. EVALUATION CRITERIA
The evaluation tests identification separately on each of the 12 most common turns, using balanced sets of the most frequent drivers. It measures accuracy with stratified, reshuffled cross-validation and includes non-ideal driving sessions.
IV. EVALUATION CRITERIA
- Each of the 12 turns is evaluated independently, using the n most frequent drivers with n varied from 2 to 5.Training sets are balanced by driver, and excess sessions are dropped in chronological order.Source: §IV · pp. 4–5
- Table I presents prediction accuracy for each of the top 12 turns.The supplied table caption identifies its scope but provides no accuracy values.Source: Table I · p. 5
- Stratified k-fold cross-validation with reshuffling measures identification accuracy, with k set to the number of sessions per driver for that turn.Sessions affected by conditions such as pedestrians or slow traffic are retained to test robustness beyond ideal turns.Source: §IV · p. 5
V. EMPIRICAL RESULTS
Single-turn identification achieved useful average accuracy, but performance varied substantially by turn and driver. Rural turns were especially strong, while the most informative sensors differed by context and the random forest compared favorably with alternatives as the candidate pool grew.
V. EMPIRICAL RESULTS
- 76.9% average accuracy for two-driver classification and 50.1% for five-driver classification, using only 8–10 seconds of sensor data.Across turns, two-driver results ranged from 55% to 93.5%, and five-driver results ranged from 33.2% to 70%.Source: §V · p. 5
- Turn #8 had the lowest five-driver accuracy; drivers 3 and 5 were identified at 2% and 8%, while drivers 1, 2, and 4 reached 60%, 57%, and 39%.The confusion matrix also shows misclassification between drivers 4 and 5; turn #11, whose five most frequent drivers differ from turn #8’s, performed strongly for all five.Source: §V · p. 5, Figure 6 · p. 5, Figure 7
- The three best-performing turns by a large margin—turns 10, 11, and 12—were rural, where fewer obstacles may make driving patterns more consistent across sessions.The paper presents reduced external interference as one explanation for better recognition of driver styles.Source: §V · p. 5
- Sensor importance varied across turns, with no single universal indicator of identity; forward acceleration, brake pedal, and steering-wheel velocity were especially relevant on the most accurate turns.The authors suggest that more unified models are needed to identify drivers holistically across turn contexts.Source: §V · pp. 5–6
- Multinomial logistic regression and SVM were comparable to the random forest for two-driver classification but fell behind as the candidate pool grew.The comparison used the same feature set.Source: §V · p. 6
VI. CONCLUSION AND FUTURE WORK
The paper applies a random-forest method to identify drivers from sensor data collected during a single turn, reporting average accuracy of 76.9% for two-driver and 50.1% for five-driver classification. It identifies longer time-series evaluation, model comparisons, and analysis across turn types as directions for future work.
VI. CONCLUSION AND FUTURE WORK
- 76.9% average accuracy for two-driver classification and 50.1% for five-driver classification across 12 frequently made turns in real-road Audi data.The dataset covers real drivers and roads around Ingolstadt, Germany.Source: §VI · p. 6
- The method uses unique patterns in individual driving habits to predict driver identity from sensor data collected during a single turn.The paper describes potential applications and extensions, including driver profiles and examining differences across turn types.Source: §VI · p. 6
- Future work includes testing the random-forest method on longer time series, comparing it with Miyajima et al., exploring additional features or sensors, and combining prediction models.The authors also propose comparing turn types to study how identifying patterns may differ across settings and to build driver profiles.Source: §VI · p. 6