Real-time SVM-based emotion recognition algorithm

The rise in ageing population is a global trend and this increasing number of elderly requires the development of new techniques, especially in the healthcare. Nowadays, already a lot of research is conducted with respect to the development of healthcare robots. However, these robots often focus on practical tasks and lack on a social interaction level. To enhance these social skills it is necessary to analyze both verbal and non-verbal communication. This paper focuses on the latter form of communication, more specific on emotion detection. To accomplish this the developed algorithm extracts specific facial cues, in the form of displacement ratios, and interprets these cues with a cascade of SVMs. In total there are 4 different steps to achieve the emotion detection. First, the countenance is detected with an adapted Histogram of Oriented Gradients (HoG) algorithm. Subsequently, 19 feature points are derived from the facial region. The next step comprises the calculation of 12 displacement ratios based on the distance between those feature points in successive frames. Finally, the displacement ratios are used as feature vectors for a multi-class SVM in cascade with a binary SVM. The developed algorithm is evaluated on the Extended Cohn-Kanade (CK+) dataset and has an overall accuracy of 89.78% with a detection speed of less than 30ms, which makes it suitable for real-time applications.