Facial expression recognition based on emotional artificial intelligence for tele-rehabilitation.

Tele-rehabilitation aims at increasing clinical outcomes while reducing costs and improving patients' quality of life (QoL). However, two main challenges need to be addressed to ensure its effectiveness: remote motor and cognitive rehabilitation. In this research work, we want to focus on the latter. Our idea is to integrate the concept of Emotional AI into tele-rehabilitation by monitoring the facial expressions of patients during motor rehabilitation exercises. Thus, we can assess the patient's cognitive and emotional state, with the objective of determining the relationship between motor and cognitive rehabilitation outcomes. Therefore, this study considers a Cloud/Edge continuum tele-rehabilitation scenario where a Hospital Cloud interacts with remote rehabilitation and monitoring Edge devices placed in patients' homes and/or rehabilitation centres. Specifically, we want to assess the performance of a Facial Expression Recognition (FER) system that can be deployed at the Edge. To achieve our goal, we employ the MediaPipe suite of libraries, which is optimized to run on low-resource devices. In particular, we used its Face Mesh module that is capable of generating a face mesh (i.e., a set of 3D face points) from an input image. The features of the mesh are then used to train a classifier that can identify the different facial expressions defined in Ekman's model (i.e., angry, fear, happy, sad, surprise, and neutral). In our experiments, we tested several combinations of datasets, face meshes (FMs), face feature maps (FFMs), and classifiers to identify the best-performing solution and demonstrate the applicability of this approach in a tele-rehabilitation environment. • Tele-rehabilitation (TR) aims at increasing clinical outcomes and patients' QoL. • Applying the concept of Emotional AI to tele-rehabilitation in a Cloud/Edge scenario. • Assessing the performance of a lightweight Facial Expression Recognition (FER) model. • MediaPipe was adopted to generate a face feature map used to train a classifier. • Experiments show the applicability of this model in a tele-rehabilitation framework. [ABSTRACT FROM AUTHOR]