Optimized, robust, real-time emotion prediction for human-robot interactions using deep learning.

To enable humanoid robots to share the social space,development in technology is required for natural interaction with the robots using multiple modes of communication such as speech, gestures, and share emotions with them. This research is targeted towards addressing the core issue of emotion recognition problem, which would require fewer computation resources and a much lesser number of network parameters, which will be more adaptive to compute on social robots for real-time communication. Any robots will have limited computation capability for run time actions and decisions. In the present investigation, Inception based Convolution Neural Network(CNN) Architecture is proposed to improve the emotion prediction. The proposed model has achieved improved accuracy of up to 6% improvement over the existing network architecture for emotion classification. The model was tested over seven different datasets to verify its robustness. In addition, real-time implementation capability is verified on humanoid robot NAO, which depicts its social behavior in real-time. The proposed model is reducing the trainable parameters to the extent of 94% as compared to vanilla CNN model, which indicates that its implementation ability in a real-time based application such as human-robot interaction. Rigorous experiments have been performed to validate the methodology, which is sufficiently robust and could achieve a high level of accuracy. Seven datasets are used to build a robust model. Finally, the model is integrated in a humanoid robot, NAO, in real-time. When averaged over all the emotions, the reduction in response time by 60% and 61% and improvement in prediction rate by 42% and 21% when compared in real-time environment with Vanilla CNN and state of the art model respectively. [ABSTRACT FROM AUTHOR]