Automatic Accuracy versus Complexity Characterization for Embedded Emotion-Sensing Platforms in Healthcare Applications

In this paper, we present a novel framework to realize an automatic “accuracy versus complexity” characterization for embeddable emotion recognition systems in healthcare applications. This framework is based on a series of grid searches able to identify highly descriptive features from the input signals, while taking into account implementation ease, memory usage, and classification performance. This paper is articulated on a proof of concept, in which the proposed framework is used to extract an emotion recognition architecture able to discriminate up to 8 emotions, by employing the users' cortical activity and a 3D model for the emotion characterization. The implemented series of grid searches lead to a final low-memory, low-resources and low-complexity emotions recognition processing chain suitable for embedded platforms. The processing chain extracted by the proposed framework has been implemented as a real-time task on a personal care robot processing core. The extracted emotion recognition system tested on cortical signals from a publicly available dataset, showed an accuracy of ~75 % (average) in discriminating 8 emotions and a memory reduction of 70% from a canonical implementation of the same feature extraction step.