Affective Recognition Based On GSR Signal By Curve Fitting And ABC Algorithm

The affective recognition based on physiological signals is an important content in the research of affective computing. And the Galvanic Skin Response (GSR), one kind of physiological signals, is an important physiological signal indicator. Researchers in the field of affective recognition generally agree that affective arousal levels can cause obvious change of GSR.On the basis of predecessors’ research, this paper studies the instantaneous variations of GSR for feature extraction, and analyzes the GSR signals under happy, angry, grief and fear four affective states using curve fitting and ABC colony algorithm. The specific work is as follows:(1) Design the affective physiological reaction experiment for the acquisition of affective physiological signal. In order to stimulate the four target affective states, experimental design team select four affective materials from more than30related film clips and video clips after repeated testing and comparison, then perfected the whole experimental procedure, and use the multi-channel physiological signal recorder MP150provided by the U.S. BIOPAC company to collect the GSR signals of people.(2) The data location of those physiological signals which contain reliable affective physiological responses. In the analysis of original data, the reliability and accuracy of affective recognition is largely dependent on the accurate selection of target GSR signal fragments. Aiming at this problem, this paper combines some methods, such as the participants report, experimental video information, affective experience records, the marked physiological signals and data cross-correlation analysis, to choose the GSR signals in an accurate time interval which has obvious affective arousal as the object of study.(3) To extract features for affective recognition from GSR signals by the method curve fitting. The GSR signal to be fitted displays obvious peaks and troughs, showing typically nonlinear features. Taking reference to the four-parameter sigmoid-exponential SCR model proposed by Lim et al., at the same time, according to the fitting effect comparison of various methods in MATLAB, this paper selected the compound function of two nonlinear function (power function and exponential function) to determine the model of custom fitting function. The function model is used to fit the rising edge of GSR signal to acquire fitting parameters, then this paper calculate two derivative variables based on the fitting parameter, use the fitting parameters and derivative variables as emotional features for "one to one" and "one to many" identification of the four target affective states. The correct identification rate is generally superior to the affective recognition based on traditional statistical characteristics of GSR signals.(4) The feature selection and classification of GSR signals based on the improved artificial bee colony algorithm. This paper uses30traditional statistical characteristics and5curve fitting parameters as original feature set together, and uses the method combining artificial bee colony algorithm and SVM classifier for binary classification recognition of happy, angry, grief and fear four affective states. The artificial bee colony algorithm introduces directional search strategy and tabu search method. Compared with the affective recognition based on single traditional statistical characteristics or curve fitting parameters, the binary classification recognition here obtaines better recognition effect.The experimental results show that the method using curve fitting function model to fit the rising edge of GSR physiological signal to extract its emotional features for affective recognition is feasible. When the mixed feature set which is made up of traditional statistical features and curve fitting parameters together is used for feature selection and classification, the curve fitting features will be choosed for the classification recognition of target affective states with higher contribution in most cases, and the experiment will get better classification results.