Source-linked AI summary
Driving Behavior Analysis through CAN Bus Data in an Uncontrolled Environment
Umberto Fugiglando, Emanuele Massaro, Paolo Santi, Sebastiano Milardo, Kacem Abida, Rainer Stahlmann, Florian Netter, Carlo Ratti
TL;DR
The paper addresses driver-behavior characterization from CAN data in uncontrolled environments, where ground truth and manageable data volumes are limited. It combines feature-based unsupervised clustering with robustness-based subsampling evaluation, finding that the dataset can be reduced by as much as 99% without impairing clustering performance, while acknowledging limits from the available driver sample and absent ground truth.
Problem
CAN-based driver characterization must identify behavior consistently in uncontrolled environments with limited surrounding-condition knowledge while reducing communication and computational load.
Method
The paper extracts features from eight CAN signals, clusters drivers unsupervised, and evaluates clustering robustness under different subsampling strategies.
Results
99% dataset reduction is possible with a suitable subsampling strategy without impairing clustering performance.
Takeaways & Limitations
The methodology supports driver-similarity analysis from CAN data collected in completely uncontrolled experiments and can support lower-data application settings.
Takeaways & Limitations
The 53-user homogeneous dataset may not contain enough behavioral variety to identify common attitudes and outliers comprehensively.
Abstract
from arXiv · showhide
Cars can nowadays record several thousands of signals through the CAN bus technology and potentially provide real-time information on the car, the driver and the surrounding environment. This paper proposes a new method for the analysis and classification of driver behavior using a selected subset of CAN bus signals, specifically gas pedal position, brake pedal pressure, steering wheel angle, steering wheel momentum, velocity, RPM, frontal and lateral acceleration. Data has been collected in a completely uncontrolled experiment, where 64 people drove 10 cars for or a total of over 2000 driving trips without any type of pre-determined driving instruction on a wide variety of road scenarios. We propose an unsupervised learning technique that clusters drivers in different groups, and offers a validation method to test the robustness of clustering in a wide range of experimental settings. The minimal amount of data needed to preserve robust driver clustering is also computed. The presented study provides a new methodology for near-real-time classification of driver behavior in uncontrolled environments.
I. INTRODUCTION
The paper investigates whether CAN bus signals can support driver-behavior analysis in uncontrolled settings while reducing the communication and computation required for real-time use. It extracts signal features and evaluates clustering-based driver similarities against a background of prior controlled or differently instrumented studies.
- Motivation: CAN bus data provide extensive information about the vehicle, driver, and surrounding environment for studying driving behavior and related mobility problems.The technology exposes data from thousands of signals and can support analyses of congestion, energy consumption, emissions, and driver habits.
- Research problem: Uncontrolled CAN-based driver characterization requires consistent behavior identification with limited knowledge of surrounding conditions and reduced communication and computational load.The paper frames these as two central challenges for real application scenarios.
- Approach: The study extracts features from CAN bus signals and assesses their usefulness for finding similarities among drivers with clustering algorithms.The feature-based approach addresses the infeasibility of processing the raw CAN output in real time.
- Related work: Prior studies used simulators, instructed maneuvers, specialized vehicles, external sensors, or small driver samples, leaving uncontrolled CAN-based driver classification insufficiently addressed.The reviewed literature includes simulator validation, instructor-directed driving, multimodal sensing, and experiments involving only two or ten drivers.
- Research gap: Existing work had not jointly addressed CAN-based driver classification in a completely uncontrolled environment and the associated communication and computational reduction problem.The paper identifies both issues as gaps in prior research.
C. Motivations
The paper’s novelty combines many drivers, completely uncontrolled driving, and a large quantity of recorded data. This setting expands CAN-based applications but limits evaluation because useful signals lack external ground-truth labels.
- Novelty: The study combines a large number of drivers, completely uncontrolled experimental settings, and a large quantity of recorded data.The authors identify this combination as the main novelty in driving-behavior analysis.
- Scope boundary: The absence of ground truth limits interpretation of signals because driver aggressiveness, driving skill, and incident history cannot serve as direct comparison labels.The limitation concerns the variety of acquired signals that lack externally verified behavioral meaning.
II. DATA COLLECTION
The dataset comes from an uncontrolled naturalistic driving experiment with 64 drivers and nearly 2,000 sessions. Analysis uses resampled CAN signals focused on eight vehicle- and driver-related measurements.
- Experimental setting: 64 drivers drove without instructions about routes, speed, or behavior, creating an uncontrolled testing condition.The experiment was conducted in Ingolstadt, Germany, using ordinary cars.
- Dataset scale: 1,987 sessions and more than 2,135 hours of driving data were recorded across 2,418 sensors.Each user drove an average of 31 sessions lasting 64 minutes.
- Signal processing: All analyzed signals were resampled at 4 Hz through linear interpolation despite the CAN bus’s nonuniform sampling.High-frequency signals were sampled at 20 Hz, while some low-frequency sensors reported only when values changed.
- Signals: The analysis concentrates on eight CAN signals: brake pressure, gas position, RPM, steering angle, steering momentum, frontal acceleration, lateral acceleration, and a vehicle-speed-related signal.The listed measurements are directly or indirectly related to driver–vehicle interaction and perceptible driving habits.
- Signal rationale: Pedal and steering signals directly reflect driver actions, whereas speed, RPM, and accelerations may reflect driving habits phenomenologically.The paper gives exceeding speed limits as an example of a habit reflected by these latter quantities.
III. GROUPING DRIVERS’ BEHAVIOR
The methodology groups drivers through feature extraction, normalization, dimensionality reduction, and unsupervised clustering. It represents each signal with seven indicators capturing values, changes, extrema, and moving-window statistics.
- Pipeline: Driver grouping consists of feature extraction, feature normalization, dimensionality reduction, and unsupervised clustering.The four-step pipeline is designed to group drivers according to common characteristics.
- Signal representation: Each signal is represented as timestamped value samples, from which seven indicators are extracted.The representation begins with signal values and acquisition timestamps.
- Feature extraction: The extracted indicators include signal values, discrete first derivatives, intervals between singular points, and local-maximum values.These features capture signal variation, peak timing, and extreme-value intensity.
- Feature overview: Table I summarizes the feature definitions, while Figure 2 illustrates the signal and selected derived features for the gas pedal.The illustrated features include the difference quotient, mean, median, and standard deviation.
- Feature extraction: Moving-window mean, median, and standard deviation are also extracted from the signals.The feature definitions use a one-minute window at the 4 Hz sampling rate.
B. Features normalization
The paper aggregates each user’s feature values across sessions, removes extreme outliers, and represents the resulting samples with normalized, shared-bin frequency histograms. Users with at least 10 hours of driving are retained for homogeneous analysis.
- Each user’s feature vector joins all sessions for that user before normalization and histogram construction.
- Values outside the 2nd–98th percentile range are removed before treating the remaining vector as statistical samples.
- Ten equal-width bins spanning the global feature range ensure that users’ histograms use the same intervals.
- Each histogram is normalized so its bin values represent the user’s distribution across the feature range.
- At least 10 hours of total driving reduces the analyzed user set from 64 to 54.
C. Dimensionality Reduction
The study applies PCA to ten-bin feature histograms before K-means segmentation, using the first two components for visualization and variance assessment. These components explain more than 80% of the variance for most signal–feature combinations, although drivers form a continuum rather than sharply separated clusters.
- K-means groups drivers according to similarities in the extracted features, while robustness and scalability require explicit assessment.
- PCA projects the ten-dimensional histograms into orthogonal components that maximize projected variance.
- Ten bins balance representation of the probability-density shape against computational feasibility for machine-learning algorithms.
- More than 80% of total variance is explained by the first two principal components for most signal–feature combinations.
- The gas-pedal Feature 1 PCA plot represents each driver as a point, with behavior appearing as a continuum rather than well-separated clusters.
D. Unsupervised Clustering
The paper uses cross-validation-inspired clustering to choose a robust number of driver groups without ground-truth labels. It evaluates robustness through repeated subsampling comparisons and finds that clustering quality depends on the signal, feature, and selected number of groups.
- Because driver attitudes and correct labels are unknown a priori, the study uses unsupervised clustering rather than supervised classification.The approach is exploratory and does not rely on ground-truth information.
- The algorithm tests K-means clusterings for K = 2 . . . 10 across 40 randomized trials.Each trial randomly permutes the feature-vector elements before producing training and validation clusterings.
- The proposed robustness criterion treats clusterings as reliable when they remain similar under subsampling of the road data.This adapts the logic of cross-validation to unsupervised clustering, where distributions arise from sampled driving data.
- The method selects the K that maximizes the V-measure, using the lowest K when values are tied.V-measure compares the similarity of clusterings generated from separate training and validation subsets.
- Feature 2 of the gas-pedal-position signal produces two groups that remain identical across all 40 repetitions, unlike K = 4.The result illustrates that some cluster counts separate drivers more robustly than others.
- Brake pressure performs best overall, followed by gas-pedal position and steering-wheel signals.The reported clustering results indicate that signal and feature choices materially affect robustness.
IV. DATASET REDUCTION
The study evaluates how to reduce CAN-bus data while preserving robust driver clustering in uncontrolled driving. Independent random subsampling outperforms contiguous subsampling, and some signal-feature combinations retain performance with 99% less data.
- Subsampling strategies: The study compares independent and contiguous subsampling strategies to assess data reduction for driver clustering.Independent subsampling is random, whereas contiguous subsampling selects consecutive elements from a circular vector.
- Evaluation design: The evaluation compares clustering from all available road data with clustering from partial data subsets.The full-data clustering serves as a reference for assessing clustering from subsampled data.
- Evaluation design: Each subsampling experiment is repeated 40 times, with K-means using the optimal cluster count for each feature.The repetitions use different random numbers and compare multiple subset sizes and both subsampling methods.
- Results: Independent subsampling performs better than contiguous subsampling across the evaluated clustering comparisons.The reported comparison covers different signals and features, with a comprehensive set of combinations provided in the appendix.
- Results: 99% data reduction is possible for some signals and features without impairing clustering performance.This corresponds to reducing the original dataset by a factor of 100 for selected combinations.
V. CONCLUSIONS
The paper develops a data-driven methodology for identifying similarities among drivers from CAN-bus data collected in completely uncontrolled conditions. It combines distributional clustering with suitable subsampling, achieving up to 99% dataset reduction without impairing clustering performance.
- Methodology: The methodology clusters drivers using seven features from eight CAN-bus signals collected in a completely uncontrolled experiment.The approach uses a distributional clustering method to delineate similarities among drivers.
- Main result: The study reports up to 99% dataset reduction without impairing clustering performance when the subsampling strategy is properly chosen.The reduction is obtained by selecting an appropriate subsampling strategy for the clustering task.
A. Discussion
The discussion frames driving behavior as a data-driven construct whose interpretation remains uncertain without external ground truth. It also identifies limited user diversity and unmodeled indicator interactions as important study boundaries.
- Interpretation: Without external ground truth, the boundary between method performance and the unpredictability of human behavior remains unclear.The authors suggest that additional field knowledge may support physical and behavioral interpretations of the resulting clusters.
- Limitations: The sample was reduced from 64 to 53 users for data homogeneity reasons, limiting the variety of behaviors available for identifying common attitudes and outliers.The authors also identify interaction among different indicators and signals within clustering as requiring further investigation.
B. Applications and future work
The proposed CAN-bus methodology is positioned for scalable, real-time driver-behavior applications. Suggested uses include risk-oriented insurance profiling, driver recognition, and integration with physical sensing technologies, subject to further study and privacy safeguards.
- Applications: CAN-bus driver characterization is presented as an application-oriented technology for wide-scale and real-time implementations.The paper links this potential to CAN-bus data being present in almost every car.
- Applications: Insurance applications could assess user accident risk from driving-session data, but further studies must compare insurer profiles with the obtained clusters.The paper notes that existing segmentation commonly relies on accident history, location, distance, and velocity.
- Applications: Driver recognition could identify a driver from CAN-bus data for security or for adapting vehicle settings to comfort or efficiency.The paper describes this as a possible driver fingerprint in an uncontrolled experimental scenario.
- Future work: Combining the modeling technique with sonar, stereo cameras, lasers, or radar could support driver-behavior modeling and safer road networks.The proposed integration is also connected to self-driving-car development.
- Privacy: Consumer-vehicle deployment requires informing customers about data use and obtaining consent for collection and sharing of personal data.The paper connects these requirements to privacy laws, customer awareness, and self-determination.