A Contextual Analysis of Driver-Facing and Dual-View Video Inputs for Distraction Detection in Naturalistic Driving Environments

Dontoh, Anthony; Ivey, Stephanie; Aboah, Armstrong · 2025 · Crossref

DOI: 10.1109/fmlds67896.2025.00059

archive: archived pipeline: cataloged verified

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

Summary

This study addresses the limitation of current distracted driving detection systems, which rely exclusively on driver-facing video and often misclassify legitimate situational awareness (e.g., checking mirrors) as distraction. The authors investigate whether incorporating road-facing video context improves detection accuracy in naturalistic driving environments. The research questions whether combining driver and road views leads to measurable performance gains and how different deep learning architectures handle this multimodal input. The methodology utilizes a dataset of 972 synchronized dual-camera recordings from real-world driving, segmented into 5-second clips. Each clip was manually labeled into six categories: hands off wheel, head turned, listening to radio, normal driving, personal grooming, and stopped. The authors benchmarked three spatiotemporal action recognition architectures—SlowFast-R50, X3D-M, and SlowOnly-R50—under two input configurations: driver-only (single-view) and stacked dual-view (driver and road views combined). Models were trained using Adam optimizer with Kinetics-400 pre-trained weights and evaluated using classification accuracy, macro F1, and weighted F1 scores. Results indicate that performance gains from dual-view inputs depend heavily on the underlying architecture. The single-pathway SlowOnly-R50 model achieved a 9.8% improvement in accuracy with dual-view inputs, suggesting its simplified temporal modeling effectively leverages additional spatial context. Conversely, the dual-pathway SlowFast-R50 model experienced a 7.2% drop in accuracy, attributed to representational conflicts between pathways. The X3D-M model performed best overall in the driver-only configuration (55.3% accuracy) but showed minimal gain with dual-view inputs. Confusion matrices revealed that dual-view inputs helped distinguish ambiguous behaviors like head turns, though some misclassification between grooming and radio interaction persisted. The study concludes that simply adding visual context is insufficient for improving distraction detection; naive input stacking can cause interference in complex architectures. The findings underscore the need for fusion-aware designs specifically engineered to integrate multi-view data. This research provides critical insights for developing robust, context-aware driver monitoring systems that reduce false positives by distinguishing between actual distractions and appropriate situational awareness.

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_with_issues.

Topics

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

• distraction detection algorithms
• visual
• external distraction
• gaze based attention detection
• visual manual
• temporal

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, theory or model