| Service robots are one important research direction in the robot industry.In recent years,service robots for education,medical,home and other industries have emerged in the market.However,the current service robot industry generally do not get a good response from users,the main reason is the lack of technical accumulation,so that many technical problems have not yet been resolved.For example,in order to interact with human friendly,the service robot’s sensor generally need to align the face actively,to achieve the effect of face to face communication.However,the existing method can just realize that someone has appeared only after the human face inter into the robot’s vision,it cannot judge through other parts of the human body.In addition,the service robot’s strangers alarm and the function of personalized communication for different users,both need the robot to carry face recognition,which requires the robot to perceive face position initiatively.It not only in a sense enhances the interaction experience between people and robots,but also correct,ly identify the different people’s identity.In the existing technical scheme,people need to cooperate with the robot to move autonomously into the robot ’ s view.However,in most cases,the person just enters into the robot’s view,the robot is difficult to perceive the complete human face information.How to let the robot perceive human face actively,and get more useful information by adjusting their own posture,is the problem that the subject focuses on solving.In response to the questions raised above,we hope to obtain a decision-making network that,can infer the target location and perform the relevant actions based on the information obtained by the current robot.This process is defined as an active face perception problem.Active face perception can be modeled as a Markov decision process,which uses a depth-enhanced learning algorithm to train agents to perform active face-sensing tasks.Based on the depth reinforce-ment learning algorithm,this paper achieves the initiative face perception,and the completion of the work can be divided into the following aspects:Firstly,the research status of active face perception is introduced,and by analyzing the domestic and foreign research results,combined with the actual situation of service robots,the problems solved in this paper are summed up:How can we find the human face automatically before the human face enters the robot vision field.Based on the theoretical study of intensive learning,deep learning and deep reinforcement learning,this paper presents a method of using active depth recognition algorithm to carry out active face perception.Secondly,this paper introduces the deep learning and reinforcement learning,which are the basic algorithm of depth reinforcement learning.Based on the summary of these algorithms’ principle,advantages and disadvantages,this paper summarizes the advantages of deep reinforcement learning in solving active face perception problem.Thirdly,the problem of active face perception is modeled as a Markov process.Theoretically,the deep reinforcement learning method which proposed in this paper can train the policy network to solve the problem of face perception.Fourthly,since the algorithm will cost a lot of human and material resources in the actual scene,so a simulation experiment scene is built.In the simulation scenario,the algorithm is verified to ensure the feasibility of the algorithm,and then the algorithm is transplanted to the actual scene to verify.Fifthly,in this paper,the depth reinforcement learning algorithm is applied to the actual scene.An active camera is used to simulate the service robot,and the robot can actively search for a human face when it sees only the human foot.Through the experimental verification,the method proposed in this paper effectively solves the problem of active face perception,through the simulation and off-line two kinds of training methods,effectively promoted the active face-sensing technology in the actual application of the scene. |
