Summary: | 碩士 === 國立成功大學 === 電腦與通信工程研究所 === 96 === For building a subject independent emotional recognition system, in this paper, we try to find out the physiological features that can best distinguish people's emotion. Different from other research, we proposed new features, and found out the emotional elicited intervals in physiological time series.
We gather physiological response signals from laboratory experiment through giving subjects six classes of emotional stimulus using movie-clips watch. Shaver’s primitive emotions were adopted for classification included love, sadness, joy, surprised, anger and fear. On analyzing of physiological signals, we adopt 4 indicators that are respiration, Blood Volume Pulse, skin temperature, and EMG.
According to the psychologist James' question, our study is based on the hypothesis: common features exist across individuals. But the difference of physiological responses is great between subjects; even the time points and the degree of experiencing emotion are inconsistent. In order to discover the best combination of features for distinguishing emotions, we find out the feature vector of the elicited part in the physiological time series at first and then select features and validate, mainly include four successive steps. Firstly, we adopt the respiration signal as the standard of physiology segmentation and propose a segmentation algorithm basis on evaluating the changing amount of the signal. Second, implement features through statistic techniques, physical motivation, and model method; and further normalize. Next, using K-means to mining out the elicited partitions according to clustering phenomenon of common features in the space. Finally, join the linear discriminate analysis in the course of Sequential search strategies for feature selection. The result shows, best features are the Linear Predictive Coefficients of respiration power spectrum, the proposed features make the classification rate of the inside data cross-validation more than 86% and up to 73.5% in testing the outside data.
|