Research On Active Person Perception And Localization Based On DQN

Service robot technology is at the forefront of modern technology research.With the rapid progress in areas such as artificial intelligence,machine learning,the development of human-machine interaction technologies for service robot is facing an important technical bottleneck.In other words,how to identify users' identities,behaviors and intentions accurately,effectively,and securely is an important technical foundation for improving the intelligence level of service robot.The primary task is to enable robot to perception and localization person.In order to further improve the problem of poor initiative and intelligence for the existing person perception technology,and on the basis of theories such as robot technology,machine learning,a person perception and localization system model is established based on Markov process and an active person perception and localization based on DQN(Deep Q-Network)algorithm is designed.The algorithm is applied in simulation and performs well.The main work of this paper includes:(1)The present situation has been analyzed and the problem has been formed.The current situation of perception and localization has been researched,and the current situation of vision based person perception and localization technology has been mainly studied.In view of the problems of poor initiative and poor environmental adaptability of traditional visual perception and localization,this paper has applied the deep reinforcement learning algorithm to enhance the initiative and flexibility of person perception and localization through training of the agent.(2)A person perception and localization system model has been established.The problem of active person perception and localization has been modeled as a Markov decision process according to the state property of the system.The discrete action space of the agent has been designed and further decomposed into a set of action in eight directions.The standard of person perception and localization under the state of target was studied,the detector model has been designed and optimized,and the return function of the system has been designed to clarify the learning goal of the agent.(3)The DQN algorithm for active person perception and localization has been achieved.A decision neural network for active person perception and localization has been designed,and the abstract expression of network abstraction has been decomposed into two parts:state value flow and action superiority flow,to ensure the agent learn more about the state of the environment.The optimization model algorithm of the active perception has been designed based on the DQN algorithm,and two different network parameters have been used to separate action selection from network optimization to improve the stability of network training.(4)The Gazebo simulation and verification analysis has been carried out.The environment model and agent model of active person perception and localization under Gazebo environment have been created,and ROS communication between models has been realized.The learning rate optimization experiment and rate of return,action number tracking experiment,and comparison test and robustness test analysis during the training phase have been carried out.The experimental results have showed that the DQN algorithm designed in this paper can better realize active person perception and localization.Finally,the work of this paper and the experience gained have been summarized,and the deficiencies of this paper and the problems that need further solution have been analyzed.