Authors:
Hany Ferdinando
1
and
Esko Alasaarela
2
Affiliations:
1
University of Oulu and Petra Christian University, Finland
;
2
University of Oulu, Finland
Keyword(s):
Emotion Recognition, Feature Fusion, NCA, ECG, EDA.
Related
Ontology
Subjects/Areas/Topics:
Applications
;
Cardiovascular Imaging and Cardiography
;
Cardiovascular Technologies
;
Health Engineering and Technology Applications
;
Pattern Recognition
;
Signal Processing
;
Software Engineering
Abstract:
Feature fusion is a common approach to improve the accuracy of the system. Several attemps have been made
using this approach on the Mahnob-HCI database for affective recognition, achieving 76% and 68% for valence
and arousal respectively as the highest achievements. This study aimed to improve the baselines for both
valence and arousal using feature fusion of HRV-based, which used the standard Heart Rate Variability analysis,
standardized to mean/standard deviation and normalized to [-1,1], and cvxEDA-based feature, calculated
based on a convex optimization approach, to get the new baselines for this database. The selected features,
after applying the sequential forward floating search (SFFS), were enhanced by the Neighborhood Component
Analysis and fed to kNN classifier to solve 3-class classification problem, validated using leave-one-out
(LOO), leave-one-subject-out (LOSO), and 10-fold cross validation methods. The standardized HRV-based
features were not selected during
the SFFS method, leaving feature fusion from normalized HRV-based and
cvxEDA-based features only. The results were compared to previous studies using both single- and multi-modality.
Applying the NCA enhanced the features such that the performances in valence set new baselines:
82.4% (LOO validation), 79.6% (10-fold cross validation), and 81.9% (LOSO validation), enhanced the best
achievement from both single- and multi-modality. For arousal, the performances were 78.3%, 78.7%, and
77.7% for LOO, LOSO, and 10-fold cross validations respectively. They outperformed the best achievement
using feature fusion but could not enhance the performance in single-modality study using cvxEDA-based
feature. Some future works include utilizing other feature extraction methods and using more sophisticated
classifier other than the simple kNN.
(More)