On the Forces of Driver Distraction: Explainable Predictions for the Visual Demand of In-Vehicle Touchscreen Interactions

Ebel, Patrick; Lingenfelder, Christoph; Vogelsang, Andreas · 2023 · arXiv

archive: archived pipeline: cataloged verified

Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of predicting and explaining the visual demand of in-vehicle touchscreen interactions to mitigate driver distraction. As infotainment systems become more complex and driving automation increases, drivers are increasingly tempted to engage in secondary tasks, significantly raising crash risks, particularly when eyes are off the road for more than two seconds. Current evaluation methods rely on expensive, late-stage empirical testing or simplified linear models (e.g., Keystroke-Level Model) that lack accuracy and fail to capture non-linear dependencies between interactions and driving contexts. The authors propose a machine learning approach that predicts visual demand based on naturalistic driving data and provides explainable insights into which design elements and driving conditions contribute to distraction, enabling safer design decisions in early development stages. The study utilizes large-scale naturalistic data collected from over 100 Mercedes-Benz test vehicles equipped with eye-tracking cameras and telematics logging. Data was gathered from 3,046 trips across various models, capturing touchscreen interactions, driving parameters (speed, steering angle, automation status), and gaze data. The researchers processed this data to define "secondary task engagements," filtering out sequences with passenger interference or missing data to yield a final dataset of 12,142 engagements. They employed machine learning models, including Random Forests, Gradient Boosting Trees, and Feedforward Neural Networks, to solve two tasks: classifying whether a "long glance" (>2 seconds) occurs and regressing the Total Glance Duration (TGD). To ensure interpretability, they applied the SHapley Additive exPlanations (SHAP) method to analyze feature contributions. The results demonstrate that the proposed machine learning approach outperforms related work in accuracy. The model identified interactions resulting in long glances with 68% accuracy and predicted total glance duration with a mean error of 2.4 seconds. The SHAP-based explanations revealed significant factors influencing visual demand, replicating findings from prior studies regarding the impact of specific UI elements, vehicle speed, and driving automation levels. For instance, the analysis highlighted how different interaction types and driving contexts non-linearly affect driver attention allocation. The data distribution analysis confirmed that the collected naturalistic data was representative of real-world driving behavior, showing distinct patterns in glance duration compared to older or more controlled datasets. The significance of this work lies in providing an automated, explainable tool for automotive UX designers to evaluate the safety of in-vehicle interfaces without requiring functional prototypes or expensive user studies. By identifying which design decisions lead to high visual demand, the method supports the creation of "safe by design" systems. The integration of explainable AI allows designers to understand not just *that* a design is distracting, but *why*, facilitating informed adjustments to UI elements and interaction flows. This approach bridges the gap between high-accuracy machine learning predictions and the interpretability required for practical design application, offering a scalable solution for assessing driver distraction in the era of increasingly automated vehicles.

Key finding

Models predicted long-glance occurrence with up to ~68% accuracy and total glance duration with a mean error of about 2.4 s, outperforming related work; SHAP explanations replicated prior findings that vehicle speed, driving automation level, and specific UI elements modulate driver visual attention allocation, demonstrating that naturalistic-data + explainable ML can flag distracting touchscreen design patterns at early design stages without simulator studies.

Methodology

naturalistic

Sample size: 12,142 secondary-task engagements from 3,046 trips (production Mercedes-Benz fleet)

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via discover_arxiv on 2026-05-04 (3 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• temporal
• visual
• visual manual
• distraction detection algorithms
• attention allocation
• visual occlusion

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Applied Guidance: design guidelines
• Empirical Findings: behavioral performance data
• Methodological Resource: measurement protocol