A robust multimodal approach for emotion recognition

摘要: Emotion recognition is one of the latest challenges in intelligent human/computer communication. Most previous work on emotion focused extracting emotions from visual or audio information separately. A novel approach presented this paper, including both and video clips, to recognize human emotion. The Facial Animation Parameters (FAPs) compliant facial feature tracking based GASM (GPU Active Shape Model) performed generate two vector streams which represent expression speech one. To extract effective features, geodesic distance estimation, we develop an enhanced Lipschitz embedding embed high dimensional acoustic features into low space. Combined with vectors, extracted terms features. Then, a tripled Hidden Markov Model introduced perform allows state asynchrony observation sequences while preserving their natural correlation over time. experimental results show that outperforms conventional approaches for recognition.