Feature Selection and Feature Extraction for Emotion Recognition Based on Multiple Short-Time Physiological Signals

• Main Authors: Shin-Hao Dai, 戴欣浩
• Other Authors: Sung-Nien Yu
• Format: Others
• Language: zh-TW
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/78424722899130791289

碩士 === 國立中正大學 === 電機工程研究所 === 103 === In this paper, we proposed an emotion recognition system based on three short-time physiological signals. Electrocardiogram (ECG), Photoplethysmorgraphy (PPG) and Skin Impedance (SI) were used to recognize five kinds of negative emotions, including neutral (non-stimulated state), sad, stress, anger and disgust. In our study, we aimed to develop a user-independent system. This emotion recognition system was composed of data acquisition (physiological signals), feature calculation, normalization, feature selection or feature extraction, and classification. First, in the data acquisition part, 50 subjects were recruited to participate in this study, including 22 males and 28 females. By employing visual and audio stimulation, the subject emotions were induced and the signals were recorded. Second, in the feature calculation part, we calculated 7 types ECG features from wave-form and HRV sequence, 10 types PPG features from wave-form and HRV sequence and 3 types SI features from wave-form and SCR sequence. Totally, 140 features were calculated. Third, we normalized our feature set to the same level. Fourth, in the feature selection part, we performed Genetic Algorithm (GA) to select the most effective feature set to enhance accuracy. On the other hand, the feature extraction part, we compared the performance of the Principal Component Analysis (PCA), Independent Component Analysis (ICA), Linear Discriminant Analysis (LDA) and 3 modified LDA (OLDA, SLDA and RLDA) methods in reducing the feature dimensions by mapping the original data to the better subspace. Finally, we used SVM to classify emotions. And we performed leave-one-out scheme for cross validation. According to the result, the accuracy were 70.4% when using GA feature selector, 67.6% when using OLDA feature extractor, 95.2% when using OLDA feature extractor in combination with the GA feature selector.