Multimodal car driver stress recognition

In this paper we address the problem of multimodal car driver stress recognition. To this aim, four different signals are considered: heart rate (HR), breathing rate (BR), palm EDA (P-EDA), and perinasal perspitation (PER-EDA). The raw signals are windowed and for each window 21 different features, including both time-domain and frequency-domain descriptors, are extracted. The recognition problem is formulated as a stress vs no-stress binary problem, and is addressed in two different experimental setups: five-fold cross validation and leave one subject out. In both setups the extracted features are classified, both individually and concatenated, with three different classifiers (k−NN, SVM, and ANN) using them both alone and stacking their predictions. Experiments run on a publicly available database of multimodal signals acquired in a controlled experiment on a driving simulator show that the best recognition results are obtained feeding the classifiers with the concatenation of the features of all the signals considered, reaching a micro average accuracy of 77.25% and 65.09% in the two experimental setups respectively.