Research On Control Strategy Of Camera Pose Based On Deep Reinforcement Learning

With the wide application of intelligent platform in daily life and production process,more and more intelligent platforms are equipped with PTZ cameras as an important means to obtain effective visual information.In order to realize the excellent characteristics of PTZ camera,it is important to study the pose control strategy of PTZ camera.Aiming at the pose control of PTZ camera,the pose control algorithm based on traditional PID method combined with swarm intelligence algorithm,and deep reinforcement learning have been proposed in this paper.The main work of this paper is as follows.Firstly,the basic concepts of reinforcement learning and deep reinforcement learning are introduced,and the characteristics of three reinforcement learning algorithms based on value function,strategy gradient and actor-critic are compared and analyzed,as well as the characteristics of typical deep reinforcement learning model DQN algorithm.Secondly,based on the traditional PID method combined with swarm intelligence algorithm,a camera pose control algorithm through the improved artificial fish swarm algorithm and adaptive fuzzy PID is presented.The technical route and constraint area of PTZ camera pose control are designed.Aiming at the shortcomings of artificial fish swarm algorithm,genetic algorithm is introduced to optimize the algorithm,and the artificial fish swarm algorithm based on genetic idea is designed.Combined with the improved artificial fish swarm algorithm and adaptive fuzzy PID algorithm,the camera pose control experiment is conducted.Compared with the conventional PID and adaptive fuzzy PID control algorithm,the effectiveness of the proposed algorithm is verified.Finally,the PTZ camera pose control algorithm based on deep reinforcement learning is proposed.The PTZ camera pose control model is analyzed,the state space and reward function are designed according to the task requirements.The intermediate reward is composed of state reward and time reward,and the final reward of success and failure is reassigned to guide the agent to converge as soon as possible.By creating the success experience pool and failure experience pool,the adaptive priority experience replay method is proposed to improve the training effect.The performance of hyperparameters such as network structure,learning rate and successful experience extraction ratio on training results are discussed.Through scenario training and testing,the proposed improved TD3 algorithm,DDPG algorithm and conventional TD3 algorithm are compared and analyzed.The test results have shown the effectiveness of the proposed algorithm.