Analyse von Methoden zur Emotionserkennung mit Wearables

Automatic emotion recognition through physiological parameters is a promising research field in the health sciences. In particular, advancements in wearable technology and cloud computing enable continuous data collection and emotion recognition, which can aid in the early diagnosis of psychological disorders such as depression. Additionally, therapy methods can be customized/designed according to psychological well-being. Commonly used sensor data include blood volume pulse, heart rate, heart rate variability, skin conductance and skin temperature. After data filtering and (statistical) feature extraction, machine learning methods are often used for classification (Support Vector Machines, K-Nearest-Neighbor, Random-Forest...). Recently, there has also been research into deep learning methods such as convolutional neural networks. There are two specific emotion models for emotion classification, the discrete model, in which predefined emotions are analyzed and the dimensional model, in which emotions are described as a combination (vectors) of several components (dimensions). The two-dimensional model is the most common, in which one axis represents the intensity and the other represents the polarity of the emotion. This study investigates whether wearables provide a good basis for emotion recognition using the two-dimensional model. The analysis shows that wearables are promising and can provide accurate results, but data must be well prepared for the classification method. In addition, a large datasets and a homogeneously distributed group of test subjects is necessary. The accuracy heavily depends on the test subjects and emotions are very subjective. Furthermore, the presented two-dimensional model is not sufficient and an extension is proposed to differentiate between similar emotions. Finally, by comparing different works, it can be assumed that there is no single best classification method/algorithm and that the best method should be “explored” for each data set.