UnSkEm: Unobtrusive Skeletal-based Emotion Recognition for User Experience

In this paper, the proposed framework utilizes body joint movement patterns extracted from skeletal joint features from Kinect v2 sensor in order to recognize emotions. Instead of using traditional methods for feature learning such as feature clustering, we proposed two methods Mesh Distance Features and Mesh Angular features to represent highly accurate body postures. For these methods, we only considered upper body joints which were 15 in number. Recognition of human emotion is performed using Support Vector Machine (SVM) which is train with Sequential Minimal Optimization (SMO). The contribution of this paper is two-fold. Firstly it uses a limited set of skeletal joints instead of tracking whole-body joint coordinates. Secondly, it uses the proposed methods of MAD and MAF for feature extraction. The proposed framework recognizes six emotions (Anger, Happiness, Sadness, Neutral, Surprise, and Fear) over the dataset collected for evaluating the User Experience platform. The experimental results show promising higher accuracies for emotional state recognition in real-time.