Emotion Recognition Based on Electrocardiogram and Electrodermal Activity
| Autor: | Shu-feng Chen, 陳書鋒 |
|---|---|
| Rok vydání: | 2010 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 98 We proposed an emotion recognition system based on electrocardiogram (ECG) and electrodermal activity (EDA) in this study. By observing the waveform of electrocardiogram, heart rate variability, and the waveform of electrodermal activity, the system could recognize four kinds of emotions, including “Neutral”, “Happy”, “Stress”, “Sad”. These emotions were frequently observed in our daily lives, and we all knew what they feel like. So, it was much easier to elicit these emotions correctly in the lab and the emotional states could stay longer. This emotion recognition system was composed of data acquisition (physiological signals), feature extraction, normalization, feature selection, and classification. First, in data acquisition part, we intended to build an user-independent system, which means that it must be applicable to everyone (more than one person). Ten people were recruited to participate in the study, short-term emotional data were collected. The data were acquired in a silent room. Thus, the subjects wouldn’t be bothered during experiments. By employing visual and audio stimulation, the target emotions were induced and the signals recorded. Second, in the feature extraction part, we calculated time-domain and frequency-domain features from HRV, many kinds of features from ECG waveform and EDA waveform, such as wavelet-based, high order statistics, and nonlinear parameters etc., intending to recruit as many as most representative features for these emotions. Third, in the normalization part, it was necessary to standardize all the features to the same level. Fourth, in the feature selection part, we performed genetic algorithms (GA) to select the most effective features to enhance accuracy. Then, the sequential backward selection (SBS) method was applied for feature dimension reduction while still retaining the same accuracy. Finally, LIBSVM was used to classify these emotions. The accuracies of using LIBSVM to classify the data into four emotional states were first tested. The accuracies of separating one emotion from another emotion and one emotion from the other emotions were also measured. The leave-one-out scheme was employed for cross-validation. According to the results, the average accuracies of classifying four categories at one time based on ECG+EDA, ECG and EDA were 100%, 95%, and 87.5%, respectively. The data length we needed to attain these high accuracies was only 60 seconds. However, this system was capable of distinguishing four emotions we were interested. Maybe in the near future, we can make it an on-line real time system.Thus, emotional state can be distinguished in no time and the information can be used to facilitate their lives. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|