Research of Emotion Recognition Using Event-Related Potentials and Machine Learning Techniques

摘要

Developing a machine's ability to recognize emotion states is one of the hallmarks of emotional intelligence and important prerequisite for high-level human computer interaction (HCI). This paper reports on a novel methodology for the automatic assessment of emotional responses. A groupof 20 participants was presented with sets of images gathered from the International Affective Pictures System (IAPS) to obtain distinct affective arousal (low, high) and valence (negative, positive) rating levels. The oddball paradigm was adopted in this experiment to study the time-spacecharacteristics of emotion in arousal-valence space. Specific affective states were suitably induced while eventrelated potentials (ERPs) were simultaneously acquired. Differences in brain activities for emotion perception and possible neural response characteristics were investigated by theanalysis of representative electrodes and relevant components from 120 ms to 1000 ms. ERP analysis show that, 4 emotion categories produced significant different amplitudes for the P3 and LPP components, as well as significant different latencies for the P3 component. Based on these conclusions,this paper implemented an automatic multiclass arousal/valence classifier using downsampled waveform for late ERP components and P3 latencies. A good recognition accuracy (>90 percent) after 10-fold cross-validation steps for 4 emotion classes was achieved by using the Extreme LearningMachine (ELM).