Emotion Recognition Based On Peripheral Physiological Signals

Emotions refer to the psychological and physiological coordinated responses of people under the stimulation of specific scenes and are closely associated with people's psychological state and physical health.Negative emotions would affect work efficiency and quality seriously,and also interferes with high-level cognitive processes such as judgment and decision-making.Therefore,there is a crucial practical application value to identify and analyze emotional state objectively and accurately.At present,commonly used methods of emotion recognition include facial expressions observation and scale-based self-assessment.Although these two methods are simple and effective,they are both subjective and lagging.Human emotional activity is regulated by the autonomic nervous system and central nervous system,as well as the human brain's advanced cognition.Hence it is closely related to changes in physiological state.The response of physiological signals is not controlled by subjective thoughts thus can objectively reflect emotional state.Therefore,emotion recognition based on physiological signals is of great significance.This thesis studies a method for identifying three emotional states(fear,peace,and joy)based on peripheral physiological signals including electrocardiograph,photoplethysmography,and respiration.The emotions of 66 subjects are induced through audiovisual stimulation and three physiological signals are collected.A total of857 valid samples are obtained eventually.A multi-signal,multi-dimensional,multi-parameter emotional feature set is constructed from three perspectives: time domain,frequency domain,and nonlinearity domain.Using the stepwise regression method for feature selection.11 effective features are selected for the three binary classification tasks of fear-peace,fear-joy,and joy-peace.16 effective features are selected for the three-categories classification task of joy,peace,and fear.Specific correlations between effective features and emotional states are also analyzed.Taking effective features as input,this thesis constructs two emotion recognition models using support vector machine and feedforward neural network,and then optimize the recognition effect by Bayesian optimization method to select model parameters.After assessing the performance of two models through confusion matrix,accuracy,ROC,AUC,and other indicators,it is found that the performance of the feedforward neural network is better than that of the support vector machine,with recognition accuracy rates of 93.3%,90.8%,92.3%,and 81.3% for the four classification tasks,respectively.Besides,the emotion recognition method in this thesis is applied to the DEAP multi-modal standard emotion database,and recognition accuracy of the three dimensions including valence,arousal,and dominance reaches85.4%,87.8%,and 84.1%,respectively.It confirms the feasibility,reliability,and generalizability of the method proposed in this thesis.