Bio-Feedback Based Human Emotion Identification Using a Minimal Set of Physiological Signals

Abstract

The thesis presents the disclosure of a new set of features to discriminate human emotions lying on negative-high arousal and positive-high arousal valance regions of arousal-valance space using bio-feedback signals that has a very low \cost of computation" in contrast to features found so far which are extremely infeasible to be used with resource constrained platforms like mobile devices. The approach we have taken di ers from the very beginning in that we consider vectors that cover small time spans (2.5 - 5.0 seconds) which are naturally useful for real-time classi cations. This is also a work around for a short-coming of existing methods where most of them have con- sidered large time spans without more emphasize on real-time classi cation of emotions. The research also tries to improve upon current studies on the eld by using only the minimal set of equipments with contrast to infeasible set of equipment used to discriminate emotions which is not suitable for resource constrained platforms. The initial assumption was that \di erent patterns of skin responses(GSR) exist for a particular person which are patterns that do not dependent on any particular day but solely on the person. We have been successful in verifying these for anger-joy. Surprisingly our new features are able to retain the same level of accuracy and some times even higher accuracies as with the high cost features use by other researches. Furthermore we have identi ed that hate-grief has no information ly- ing in the skin response patterns that make them eligible for discrimination only with GSR which we veri ed with further testing. We have also iden- ti ed some redundancy in a high cost feature set used by MIT in a related research([12]) which was evident with self organizing maps. At a glance the high cost features maintained 76.7% and 50% for anger-joy and hate-grief respectively. But our new feature set was performing well above the ac- curacy and reached 80% mark with a new method derived to improve the accuracy.