How Facial Features Convey Attention in Stationary Environments

Domantay, Janelle; Morris, Brendan · 2022 · Crossref

DOI: 10.9741/2766-7227.1021

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 challenge of detecting user attention and fatigue in stationary environments, such as online classrooms or computer-based workspaces, moving beyond the traditional focus on driver safety. The research aims to identify which visual facial features most effectively predict awareness levels and to compare the performance of classical machine learning methods against deep learning approaches. Specifically, the paper investigates whether Support-Vector Machines (SVMs) utilizing explicit feature extraction can compete with Convolutional Neural Networks (CNNs) and Convolutional Recurrent Neural Networks (CRNNs) in terms of accuracy and computational efficiency. The methodology utilized the University of Texas at Arlington Real-Life Drowsiness Dataset (UTA-RLDD), from which a subset of 3,000 frames was selected to represent three alertness classes: Alert, Low Vigilance, and Drowsy. Visual features were extracted using the open-source toolkit OpenFace, focusing on Histogram of Oriented Gradients (HOG), Facial Action Units (AUs), and eyelid visibility metrics approximating PERCLOS. The study trained SVM classifiers using various kernels on these extracted features. In parallel, deep learning models—including MobileNetV2, ResNet50, InceptionV3, and DenseNet121—were trained both from scratch and via transfer learning on individual frames. Additionally, CRNNs and 3D CNNs were evaluated on 28-frame video sequences to capture spatio-temporal dynamics. Performance was measured by test accuracy, precision, recall, and processing time. The results indicated that HOG features were the strongest predictors of drowsiness, outperforming Action Units when used in isolation. SVMs utilizing HOG features with Polynomial or Gaussian kernels achieved high validation accuracies, with the best SVM model reaching 96.04% accuracy when analyzing video segments with both HOG and AU features. Deep learning models generally achieved higher peak accuracies, with InceptionV3 reaching approximately 99.6% validation accuracy when trained from scratch. However, the study found that CRNNs did not significantly outperform their CNN counterparts. Crucially, while deep learning methods offered superior accuracy, SVMs required substantially less computational resources and processing time, operating on a CPU in milliseconds compared to seconds for GPU-accelerated neural networks. The significance of this work lies in demonstrating that classical machine learning methods, particularly SVMs with HOG features, can approach the performance of deep learning models while offering greater explainability and lower resource requirements. This suggests that for applications where computational efficiency and feature interpretability are prioritized over marginal gains in accuracy, SVM-based approaches are viable alternatives to complex neural networks for attention detection in stationary settings.

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 2026-06-10; verification: verified.

Topics

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

• distraction detection algorithms
• drowsiness detection algorithms
• gaze based attention detection