Research Of Facial Expression Transfer For Remote Human-robot Interaction System

Facial expression is one of the most important and direct carriers of human emotion expression and communication.It has the characteristics of direct perception and strong performance.The expression transfer algorithm uses expression capture algorithm and face encoding system to copy facial expressions to a specific target,which has been widely used in the animation production industry.The traditional facial expression transfer algorithm has the problems of light sensitivity,poor generalization ability for the human face rotation and personalization.The main research contents of this thesis is as follows:For the problems of light sensitivity and poor robustness of the human face,it was noticed that the problem mainly affected the accuracy of expression capture of facial expression transfer,and the convolutional neural network always get better results on illumination and occlusion issues in other image detection tasks.Therefore,this paper proposes to use the convolutional neural network to train the deep model to obtain the deep features of facial expression.However,the convolutional neural network often has a huge parameter matrix,once the network depth is slightly larger,it is difficult to meet the real-time requirements of expression transfer.For the problem that it is difficult to meet the real-time performance,this thesis is inspired by the excellent effect of adaboost face detection.It eliminates a large number of error samples through a cascade of forward-stage simple networks,and then connects follow network to optimization detection results to accelerate the entire network running speed.in addition,the expression capture part includes two steps of face detection and face landmarks detection.The traditional method splits the two tasks,ignoring the correlation of the two tasks.This paper proposes the use of multi-task methods,in the cascade network,based on the characteristics of the face feature sharing,complete parallel face detection and face landmarks detection,and further accelerate the network speed.However,because this kind of method simplifies the network structure and satisfies the real-time requirements,the accuracy of the face landmarks is difficult to meet the requirements.In order to solve the problem of poor accuracy of face landmarks on the face,this paper proposes to locate the critical face points in the forward-stage network and locate the key points of the face,positioning 4 parts: The approximate position of the left eye,right eye,nose and mouth.This four parts is then segmented by the neighborhood image of the area and used as a 4-channel input of the further convolutional neural network to perform accurate positioning again,thereby detecting the precise position of the key point of each face.Finally,this thesis also applies the expression transfer algorithm to the remote human-robot interaction system.The traditional remote human-robot interaction system has a single feedback method and lacks humanized feedback.This thesis relies on the expressive features of facial expression transfer.It embeds facial expressions into remote human-machine systems and validates the effectiveness of the system.Finally,an interactive humanoid remote human-robot interaction system is realized.Improve the interactive experience of remote human-computer interaction.