Design And Implementation Of SLAM Parameter Adaptive Adjustment System Based On Deep Reinforcement Learning

Augmented Reality integrates virtual information into reality.It makes the virtual information consistent with reality in terms of illumination,geometry and temporality.It enhances users’ perception about the real scene.As one of the core technology of Augmented Reality,SLAM calculates the relative position of sensor pose and real scene in real time,and then accurately adds virtual information.The accurate parameter threshold is the key element to achieve precise pose using SLAM.However,many visual SLAM algorithms rely on setting empirical values,and using fixed parameter configurations in practical applications.The parameters need to be re-tuned while the scene changes.To solve this problem,the work of thesis is as follows:1.This dissertation proposes an adaptive adjustment algorithm for visual SLAM parameters based on deep recurrent Q network.Firstly,visual SLAM parameters are selected as the element of action space.Secondly,the covariance matrix of the map point or the RMSE of the pose translation is used to represent the visual localization uncertainty,which is used to construct the reward function.Finally,in order to achieve the purpose of adaptively adjusting the parameters,the parameter agent selects the action with the largest Q value through the ε-greedy strategy,and then transmits it to the visual SLAM environment.2.This dissertation designs and implements a camera pose estimation system based on parameter adaptation.The system consists of server terminal and mobile one.To realize pose estimation of the mobile terminal in the real scene real-time.The server terminal estimates the camera pose and mapping.The mobile terminal shows the obtained camera pose and map,and realizes mobile augmented reality interaction based on video sequences.In this paper,a deep recurrent Q-network is used for parameter adaptive adjustment in the visual SLAM.The experimental results show that the proposed algorithm not only improves the accuracy of pose estimation,but also obtains the corresponding parameter adjustment strategy,and makes the system work well in different scenes.Therefore,the proposed system has theoretical significance and practical value.