Full text
Efficient and Compact Face Descriptor for Driver Drowsiness Detection Abdelmalik Moujahid1, Fadi Dornaika1,2, Ignacio Arganda-Carreras1,2,3, Jorge Reta1 1University of the Basque Country (UPV/EHU), Spain 2IKERBASQUE, Basque Foundation for Science, Spain 3Donostia International Physics Center (DIPC), San Sebastian, Spain Published in: Expert Systems With Applications, Volume 168, 2021, Article 114334, Elsevier. DOI: 10.1016/j.eswa.2020.114334. Available online 30 November 2020. Abstract Current advances in driver drowsiness detection consist of a variety of innovative technologies generally based on driver state monitoring systems. Extracting effective and relevant features to characterize drowsy symptoms in images and videos is still an open topic. In this work, we introduce a face monitoring system based on a compact face texture descriptor able to cover the most discriminant drowsy features. The compactness has been achieved by both a multi-scale pyramidal face representation that capture the main characteristics of local and global information, and the feature selection process applied on the raw extracted features. The proposed framework is rolled out in four phases: (i) face detection and alignment; (ii) Pyramid-Multi Level (PML) face representation; (iii) face description using a multi-level multi scale feature extraction; and (vi) feature subset selection and classification. Experiments conducted on the public dataset NTH Drowsy Driver Detection (NTHUDDD) show the effectiveness of the proposed face descriptor and the associated selection schemes. The results show that the proposed method compares favorably with several approaches including those based on deep Convolutional Neural Networks. Main Contributions •Introduces a multi-scale pyramidal face descriptor combining local and global texture features, tailored to discriminating drowsiness-related facial expressions. •Applies dimensionality reduction and feature selection methods (PCA, Fisher score) to produce highly compact and discriminant descriptors. •Develops a robust SVM-based classification framework and a scheme for blending decisions from multiple descriptors, improving detection accuracy. •Benchmarks the approach against both classical and deep learning-based state-of-the-art methods, demonstrating competitive or superior accuracy on the NTHU Drowsy Driver Detection dataset. •Proposes a fusion strategy to combine information from HOG, Covariance, and LBP descriptors, further enhancing driver state detection in challenging scenarios. 1
Impact of the Paper This study presents an efficient computer vision solution for real-time driver drowsiness detection, a critical area for reducing traffic accidents and enhancing road safety. By extracting and selecting discriminative face features using a compact multi-scale strategy, the system provides robust performance even under varied conditions typical of real-world driving. The method’s success in matching or surpassing deep learning alternatives demonstrates the value of feature engineering and interpretable modeling for embedded automotive applications. The approach enables practical deployment in intelligent vehicle systems and offers a scalable framework easily adaptable to other fatigue-related monitoring tasks. Reference: Abdelmalik Moujahid, Fadi Dornaika, Ignacio Arganda-Carreras, Jorge Reta. “Efficient and compact face descriptor for driver drowsiness detection.” Expert Systems With Applications, Volume 168, 2021, Article 114334. DOI: 10.1016/j.eswa.2020.114334 2
