| Abstrakt: |
Objective: The goal of the project is to enable the Pepper robot, developed to assist the audiology tests of hearing-impaired children, to recognize children's emotions and to regulate the behavior and feedback process in accordance with the child's mood. In order to achieve this goal, in this study, it is aimed to recognize emotion from physiological signals with machine learning. Methods: While 9 children (6 male) ages between 4-6 were shown positive and negative emotional videos and a neutral one as control, skin conductance (SC), blood volume pulse (BAT) and temperature (TEMP) were obtained using E4 Wristband. For machine learning, the average, variance and mean of the first derivative of all three signals were determined as features. For additional features, heart rate (HR), heart rate variability (HRV) were calculated from the BAT signal, and the HRV spectrum was divided into three ranges: very low (VLF) (0-0.05 Hz), low (LF) (0.05-0.15 Hz) and high frequency (HF) (0.15-0.4 Hz). For feature preparation; metric extraction, classification and clustering preprocessing steps were applied. Machine learning algorithms have been implemented using the Weka program. While processing physiological data, the attempt was to find a system where maximum success is achieved by selecting the calculated features, algorithm and classification type in different combinations. Results: The best results (around 80%) were obtained with K-Star and Random forest algorithms using LF, HF, LF/HF, average BAT, variance of BAT and first order derivative of BAT for neutral-negative, and with SVM algorithm using LF, HF and LF/HF metrics for neutral-non neutral classifications. High results were not obtained from SC or TEMP. Conclusion: Among BAT, SC and TEMP physiological data, BAT provides the best features for emotion recognition from physiological signals. Different algorithms for different classifications (neutral-negative or neutral-non neutral) provide maximum success (80%). [ABSTRACT FROM AUTHOR] |
