Driver Distraction Detection Using Artificial Intelligence and Smart Devices

Papatheocharous, Efi; Buffoni, David; Maurer, Matthias; Wallberg, Anders; Ezquerro, Gonzalo · 2024 · OpenAlex-citations

DOI: 10.1007/978-3-031-54049-3_16

archive: archived pipeline: cataloged verified

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

Summary

This paper addresses the critical safety issue of driver distraction, specifically focusing on the role of smart devices like smartphones and smartwatches. While these devices are significant sources of distraction through calls, messages, and app usage, they also offer a unique opportunity for monitoring driver behavior due to their built-in sensors and widespread adoption. The authors propose a manufacturer-independent solution called "Smart Devices Distracted Driving Detection," which leverages Artificial Intelligence (AI) and Machine Learning (ML) to detect distraction events using data from these devices. The goal is to create a scalable, retrofit-friendly system that can improve traffic safety and provide feedback to drivers, addressing the limitations of proprietary vehicle-based systems. The study presents a conceptual framework comprising four main components: a smartphone application, a smartwatch application, a camera-based computer vision system, and a web dashboard for visualization. The first three components utilize ML models trained on sensor data. For the smartphone application, IMU (gyroscope and accelerometer) and GPS data were collected during test drives to classify distraction states (e.g., phone calls, app usage) versus attention states. Similarly, the smartwatch application collected wrist motion and GPS data to detect interactions with the watch. The computer vision component used a hybrid dataset combining the public Statefarm dataset with custom recordings to classify driving normally, smartphone usage, and smartwatch usage. Data preprocessing involved cleaning, unifying time references, and handling class imbalances, particularly for smartwatch interactions which constituted less than 5% of smartwatch data and were a minority class in the vision dataset. Experimental results demonstrated varying levels of success across the three detection methods. The LSTM-based neural network for the smartphone application achieved 94% accuracy in distinguishing between distraction and attention states, with errors primarily occurring during state transitions. The smartwatch application, also using an LSTM model with dropout layers to prevent overfitting, similarly achieved 94% accuracy on the test set. The computer vision application, however, faced challenges due to data heterogeneity and imbalance. Using transfer learning with EfficientNetB0, the model initially performed poorly. However, applying data augmentation techniques improved the overall accuracy from 54% to 72%, though the F1 score for smartwatch detection remained low at 0.48. The authors successfully optimized the computer vision model for deployment by quantizing weights, reducing the model size by a factor of four with minimal accuracy loss. The significance of this work lies in demonstrating a viable, device-agnostic approach to driver distraction detection that does not rely on proprietary vehicle hardware. By deploying ML models directly on edge devices (smartphones and smartwatches) using frameworks like TensorFlow Lite and CoreML, the system ensures privacy and scalability. The integration of these detection modules into a collaborative web dashboard allows for comprehensive session analysis. The study highlights that while smartphone and smartwatch sensor data can effectively detect device-specific distractions, computer vision approaches require more diverse and balanced data to reliably detect wearable usage. This framework provides a foundation for future developments in affordable, upgradable driver monitoring systems that can be integrated into existing vehicles.

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.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-25; verification: verified.

Topics

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

• distraction detection algorithms
• visual
• mobile phones

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).

• Methodological Resource: tool software
• Theoretical Contribution: conceptual framework