Emotion Recognition from Physiological Signals Collected with a Wrist Device and Emotional Recall

Figure 1. Overview of methods for signal processing, feature engineering, classification, and validation used in this study. The subject-independent approach is highlighted in black, while methods used for the purpose of validation and investigation of subject-dependency are depicted in grey.

Figure 2. Overview of the chosen elicited emotions in the valance-arousal planes according to Russell [21].

Figure 3. Experimental protocol for a random order of emotion excitation tasks performed for five minutes each and intermittent distraction tasks for five minutes.

Figure 4. Overview of preprocessing for all signals. The preprocessed signals are denoted with a “pp” after the name. The next step after signal processing is segmentation.

Figure 5. Overview of all preprocessed signals and segmentation. The placement of all viable segments with a length of 10 s is visualized for the inter-beat interval and blood volume pulse data with a dotted line, while the selected segment closest to the middle is marked with a solid line.

Figure 6. Corruption of multi-subject data set visualized as heatmaps. Darker colors indicate a signal that is noisy or unclear. (a) The longest segment length for all emotion tasks in seconds (the maximum value for the colorbar is set to 30). (b) The number of SCRs with an amplitude above 0.01 μS per emotion task (the maximum value for the colorbar is set to 10).

Figure 7. Predictor ranks computed with the ReliefF algorithm for the emotion label and segment length of 20s.

Figure 8. Confusion matrix for the classification for the emotion label, segment length of 20 s, 10-fold validation and a trained SVM with an accuracy of 39%. Green highlights the correctly classified samples and pink the incorrectly classified samples, where darker shades indicate a higher number of samples.