Recognizing Social Touch Gestures using Optimized Class-weighted CNN-LSTM Networks

Socially aware robotic applications such as companion and therapeutic robots usually require human emotions or intent to be conveyed. As the scope of these applications increases, the need for recognizing affective touch gestures which are often used to convey these emotions or intent becomes eminent. However, existing touch gesture recognition modalities either have low recognition accuracy or depend heavily on carefully hand-crafted features, therefore limiting their deployment in real-life applications. Motivated by the need for learning models with superior accuracy which do not rely on manually selected hand-crafted features, this paper proposes an optimized class-weighted CNN-LSTM for social touch gesture recognition evaluated on the CoST and HAART datasets. Specifically, contrary to vanilla training schemes where equal importance is given to each class in the dataset, different class weights are introduced to give priority to classes that are difficult for the network to distinguish during training. Furthermore, the weights associated with each of the classes are obtained through optimization using Genetic Algorithm. The proposed model demonstrates superior performance compared with other existing models in the literature.