Uncorrelated fuzzy neighborhood preserving analysis based feature projection for driver drowsiness recognition

  • Authors:

  • Rami N. Khushaba;Sarath Kodagoda;Sara Lal;Gamini Dissanayake

  • Affiliations:

  • Centre for Intelligent Mechatronic Systems (CIMS), Faculty of Engineering and Information Technology, University of Technology, Sydney, Broadway, NSW 2007, Australia;Centre for Intelligent Mechatronic Systems (CIMS), Faculty of Engineering and Information Technology, University of Technology, Sydney, Broadway, NSW 2007, Australia;Department of Medical and Molecular Biosciences, Faculty of Science, University of Technology, Sydney, Broadway, NSW 2007, Australia;Centre for Intelligent Mechatronic Systems (CIMS), Faculty of Engineering and Information Technology, University of Technology, Sydney, Broadway, NSW 2007, Australia

  • Venue:
  • Fuzzy Sets and Systems
  • Year:
  • 2013

Quantified Score

Hi-index 0.20

Visualization

Abstract

Driver drowsiness is reported as one of the main causal factors in many traffic accidents as it progressively impairs the driver's awareness about external events. Drowsiness detection can be approached through monitoring physiological signals while driving to correlate drowsiness with the change in the corresponding patterns of the Electroencephalogram (EEG), Electrooculogram (EOG), and Electrocardiogram (ECG) signals. The main challenge in such an approach is to extract a set of features that can highly discriminate between the different drowsiness levels. This paper proposes a new Fuzzy Neighborhood Preserving Analysis (FNPA) feature projection method that is used to extract the discriminant information relevant to the loss of attention caused by drowsiness. Unlike existing methods, FNPA considers the fuzzy memberships of the input measurements into the different classes while constructing the graph Laplacian. Thus, it is able to identify both the discriminant and the geometrical structure of the input data while accounting for the overlapping nature of the drowsiness patterns. Furthermore, in order to address the singularity problem that occurs in many real world problems, the singular value decomposition (SVD), and later the QR-Decomposition, are utilized to extract a set of statistically uncorrelated features presenting the Uncorrelated FNPA (UFNPA). In the current preliminary study with datasets collected from 31 subjects only, while performing a driving simulation task, the proposed method is capable of accurately classifying the drowsiness levels using a small number of features with an average accuracy of ~93%. On the other hand, the possibility of developing a subject-independent drowsiness recognition system is also investigated when the problem is converted into a binary classification task, as imposed by the number of drowsiness levels exhibited by the drivers, with accuracies ranging from 82%-to-84%.