SVM point-based real-time emotion detection

• Published in: 2017 IEEE Conference on Dependable and Secure Computing pp. 86 - 92
• Main Authors: Swinkels, Wout, Claesen, Luc, Feng Xiao, Haibin Shen
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.08.2017
• Subjects: Emotion detection; Ensemble of Regression Trees; Eyebrows; Feature extraction; Histograms; HoG; Mouth; Nose; Real-time; Regression tree analysis; Support vector machines; SVM
• DOI: 10.1109/DESEC.2017.8073838

Face recognition is nowadays implemented in security systems to grant access to areas that are only allowed for authorized persons. However an additional layer of security can be added to these systems by detemining if the person in front of the camera is present in real-life and that the detected object is not a 2D representation of that person. Forcing people to interact with the system by for example posing a certain emotion can be an additional layer of complexity to deny the access for unauthorized persons. This paper focuses on that aspect i.e. real-time emotion detection. Therefore a novel algorithm is developed to extract emotions based on the movement of 19 feature points. These feature points are located in different regions of the countenance such as the mouth, eyes, eyebrows and nose. To obtain the feature points an Ensemble of Regression Trees [1] is constructed. After the extraction of the feature points 12 distances, in and around these facial regions, are calculated to be used in displacement ratios. In the final step, the algorithm inputs the displacement ratios to a classification algorithm, which is a cascade of a multi-class support vector machine (SVM) and a binary SVM. Experimental results on the Extended Cohn-Kanade dataset (CK+) [2], [3] indicate that the proposed algorithm reaches an average accuracy of 89,78% at a detection speed of less than 30 ms. The accuracy is comparable with state-of-the-art emotion detection algorithms and outperforms these algorithms when detecting the emotions Contempt, Disgust, Fear and Surprise. The detection speed evaluation of the proposed algorithm was perfomed on a Windows 8.1 laptop with an Intel-Core i7-5500U CPU (2.40 GHz) and 8,00GB of RAM.