A Real-Time Solution Enabling Humanoids to Efficiently Identify Faces and Facial Expressions

Owing to the advantages and effectiveness of using humanoids in the field of therapy and rehabilitation, there is a need for robots to have the capability to recognize a person and understand his/ her emotional state based on facial expressions, thus making the human-robot interaction more natural. In this thesis, an accurate, real-time and power efficient solution for face recognition and facial expression recognition is presented. The solution consists of a combination of a convolutional neural network (CNN) and a Support Vector Machine (SVM), which is deployed on NVIDIA Jetson TX2, a cheap, powerful and small sized hardware processing platform. For efficient deployment, a study on power consumption and performance of standard deep learning networks is drawn to analyze and find out the best hardware configuration of NVIDIA Jetson TX2 for inferring networks. The proposed solution was compared with AlexNet [Krizhevsky, Sutskever, and Hinson, Advances in Neural Information Processing Systems, 1097-1105 (2012)] and was found to be more accurate on facial expression datasets considered. It also has smaller model size, faster inference, lesser number of trainable parameters and consumes lesser power. The performance and functionality of the developed application was tested on videos and humans in a real Human-Robot Interaction scenario. The results were satisfactory, vindicating the fact that the application can be deployed and used in the real world.