The Research On Affective Recognition Based On Physiological Signals

Intelligent human-computer interaction has always been the research focus of researchers in artificial intelligence and other fields.Emotion recognition,as a bridge connecting humans and computers,plays a vital role in it.Previous researchers mostly used facial expressions and voice signals to perform emotion recognition tasks,but these data are subject to subjective control of people and cannot truly reflect the true emotional state of people.Emotion recognition based on physiological signals overcomes this shortcoming.It has better robustness and objectivity.In addition,different physiological signals carry more abundant emotional information,which provides an important clue for more accurate recognition of human emotions.This thesis focuses on the relationship between different emotions and physiological signals.The main contents are as follows:1.We designed an emotion induction experiment,through audio-visual materials,induced 52 subjects of different emotions,then divided emotion into low valence?high valence(LV?HV)and low arousal?high arousal(LA?HA).Simultaneously,we used the MP150 physiological signal recorder to collect electrocardiogram(ECG)?Skin electrical activity(EDA)and Respiration(RSP)in resting state and stimulating state,then the professional completed the labeling of the emotional data in conjunction with the subject's self-assessment report.The corresponding physiological features were extracted from the labeled database,and the database was evaluated by three traditional machine learning methods.The results were used as base results of the database.For valence,the best result is the classification accuracy is 60.3±0.8%and the f1-score is 60.0±1.0%.For arousal,the best result is the classification accuracy is 62.9±0.9% and the f1-score is 62.8±1.0%.Simultaneously,to judge our affective database,we compare our database with three public databases(DEAP,DREAMER and AMIGOS).2.In view of different physiological signals,our research presents a multi-view deep learning model for processing multimodal physiological signals.This model based on 1-dimensional convolutional neural network and can automatically complete the feature extraction of different physiological signals according to the characteristics of the signal by setting different parameters in different channels.Our proposed model got a classification accuracy of 67.1±1.2% and a f1-score of 66.6±1.2% in valence,and a classification accuracy of 78.1±1.4% and a f1-score of 78.0±1.4% in arousal,which has a tremendous improvement compared to traditional machine learning algorithms.To verify the validity of our proposed model,we also carried out experiments on two public databases(DEAP and DREAMER),and our proposed method had achieved a more ideal recognition result.