Emotion Recognition from Body Movements with AS-LSTM

Abstract

With the development of artificial intelligence, people's demand for emotional interaction in virtual reality experience is becoming higher and higher. When the traditional emotion recognition method is used in virtual reality emotion recognition, it has some problems, such as tedious wearing, high demand on image clarity, inaccurate emotion recognition in motion, etc. Therefore, we propose a stack LSTM network based on attention (AS-LSTM) for emotion recognition from whole body movements in VR environment. According to the importance degree, different attention values are set for the feature sequences data of each joint point in the frame sequence of human body motion by adding attention mechanism to the basis of traditional LSTM network, which will build a particular distribution of attention, focus on the key joint points affecting emotion recognition, and reduce the invalid information. Then the proposed method can improve the learning ability of network and emotion recognition accuracy. Moreover, the equipment is simple and easy to operate, which provides users a more immersive emotional interaction experience. One can observe, this method achieves higher recognition accuracy on classification of seven kinds of emotions (happy, sad, fear, anger, surprised and disgust) compared with other deep learning methods in VR. In addition, the accuracy of emotion recognition in six categories (happy, sad, fear, anger, surprise, disgust) and four categories (happy, sad, fear, and anger) is also improved.