Five-level Drowsiness Estimation Using BlendShape Features Captured by a Smartphone’s Front-facing Camera

Suzuki, Shunki; Tanaka, Hisaya · 2025 · Crossref

DOI: 10.54941/ahfe1006877

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 need for objective, automated drowsiness detection to mitigate risks associated with traffic accidents, workplace errors, and decreased academic performance. While prior research has primarily focused on eye and mouth cues using libraries like MediaPipe, this work explores the use of Apple’s ARKit to capture 52 BlendShape features and 3D head orientation via a smartphone’s front-facing camera. The authors propose a five-level drowsiness estimation method, aiming to leverage the accessibility of mobile devices for real-time, non-contact monitoring. The study specifically investigates whether ARKit-derived indicators can support multi-level classification and evaluates the relative importance of various facial features in predicting drowsiness. The experimental design involved six participants performing a 30-minute simulated driving task using Euro Truck Simulator 2 to induce drowsiness. Facial data were captured at 60 Hz using a custom iOS application, extracting spatial and temporal features such as mean values, standard deviations, and angular velocities over 5-second intervals. These 250 indicators served as inputs for a K-nearest neighbor (KNN) classifier. Ground truth labels were established by external raters who evaluated video segments using a five-point scale based on the Kitajima et al. (1997) system. To assess robustness, the data were also categorized into binary and ternary classifications. Model performance was evaluated using accuracy, recall, precision, and F1 scores, with SHapley Additive exPlanations (SHAP) analysis employed to interpret feature contributions. The results demonstrated high classification accuracy across all levels. The binary classification achieved an F1 score of 99.28%, the ternary classification reached 93.94%, and the five-class classification attained 73.10%. SHAP analysis revealed that temporal variations, particularly in mouth and head movements, were more predictive than static spatial features. Specifically, indicators related to the mouth corners and vertical head sway contributed significantly to drowsiness predictions, aligning with rater observations of yawning and head nodding. The study found that variance in eye closure rates and head sway was a crucial differentiator, as static mean values showed less distinction between drowsiness levels. The lower recall for the five-class model was attributed to a scarcity of data points for the highest drowsiness level. The findings confirm that smartphones equipped with ARKit are viable tools for accurate, multi-level drowsiness estimation. By utilizing variance and movement-based indicators rather than static spatial data, the system effectively captures the dynamic nature of fatigue. This approach supports the development of accessible, non-contact monitoring systems for driver safety and behavioral assessment. The authors conclude that future improvements could involve optimizing feature selection and incorporating additional indicators, such as hand movements, to further enhance detection accuracy.

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.

StageOutcomeToolModelPromptAttemptsCompleted
discover success Crossref 1 2026-08-09
archive success canonical_url 1 2026-08-09
extract success pdftotext 4 2026-08-10
clean success clean 2 2026-08-10
chunk success chunk 2 2026-08-10
embed success embed Qwen/Qwen3-Embedding-8B 2 2026-08-10
promote success 1 2026-08-09
summarize success llm qwen3.6-27b-nvidia summ-v5 2 2026-08-10
tag success vector_similarity 17 2026-08-11
verify success 1 2026-08-10

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

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

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).