Texting and Driving Recognition leveraging the Front Camera of Smartphones

Montori, Federico; Spallone, Marco; Bedogni, Luca · 2023 · Crossref

DOI: 10.1109/ccnc51644.2023.10060838

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 challenge of recognizing whether a smartphone user is driving or acting as a passenger, a critical issue for road safety and context-aware computing. While distracted driving causes significant injuries and fatalities, existing solutions often rely on inertial sensors or voluntary third-party applications, which are prone to error, easily bypassed by users, or lack system-level integration. The authors propose a novel approach leveraging the smartphone’s front-facing camera to classify the user’s position within the vehicle. By determining if the user is in the driver’s or passenger’s seat, the system can automatically restrict smartphone usage while driving, enhancing safety without requiring additional hardware or user intervention. The study employs a computer vision pipeline using Convolutional Neural Networks (CNNs) and object detection. Initially, the authors attempted instantaneous binary classification using standalone images, training a custom CNN and a ResNet50 model via transfer learning on a dataset of 50 augmented images. These methods yielded poor accuracy, ranging from 52.3% to 56.7%, due to the inability of generic models to focus on discriminative features amidst visual noise. Consequently, the approach shifted to object detection using the COCOv6n model. The system identifies four specific objects: the driver’s and passenger’s windows, and the respective security belts. Classification is determined by counting detected objects; for instance, more driver-related objects indicate the user is driving. This method improved standalone accuracy to 68.7%, though nearly half of the predictions remained uncertain. To resolve uncertainty, the authors developed a continuous classification system using short video sequences. They created a "Continuous dataset" comprising 1,200 frames from eight 10-second videos recorded at 15 frames per second. The system analyzes sliding windows of frames, aggregating object detections to reach a confidence threshold. Results demonstrated that accuracy increases with window size, reaching 88.3% for windows of 10 frames. Crucially, the system achieves high confidence rapidly; for a 99% confidence threshold, the optimal window size of 15 frames requires only approximately 1.4 to 2.8 seconds of video footage. This confirms that the method can reliably distinguish between driver and passenger positions within a few seconds, significantly outperforming previous inertial sensor-based approaches in both speed and accuracy. The significance of this work lies in its potential for integration into operating systems or automotive black boxes, providing a robust, hardware-free solution to mitigate distracted driving. By utilizing the ubiquitous smartphone front camera, the proposed system offers a practical path toward mandatory safety measures, such as those mandated by the European Union for vehicle data recording. The findings demonstrate that vision-based recognition of seating position is feasible and efficient, paving the way for context-aware applications that can automatically disable distracting functions when a user is identified as the driver, thereby improving overall road safety.

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 2026-06-26; verification: verified.

Topics

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

• distraction detection algorithms
• 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).

• Empirical Findings: observational prevalence
• Methodological Resource: tool software
• Theoretical Contribution: computational model