Application Research Of Start To Target Autonomous Control Of Virtual Camera Based On Reinforcement Learning

In the virtual museum scene created using the Unity3 D development engine,to realize the virtual camera’s self-display of the exhibits in the virtual scene,the existing method is to manually set the track,make the virtual camera move along the track and display the exhibits along the way,which requires a lot of manpower to adjust the key points of the track in detail.Acommon path planning method is to calculate the optimal path through algorithms given a well-defined map and starting andending points.In the virtual scene,the display of exhibits needs to consider the display angle of the virtual camera,which cannot be solved by the path planning method.In terms of the above problems,this thesis proposes a new autonomous control method to reduce human work and realize the display of exhibits.On the basis of not obtaining the position information of the exhibits in the scene,the target detection module provides the status information of the detected objects,passes it to the reinforcement learning module for training,and finally the reinforcement learning module realizes the autonomous control of the virtual camera from the start to the target to complete the display of the exhibits in the scene.And research from the following points.(1)Astate space suitable for this research topic is designed,which provides state information support forthedesignofrewardfunctions.(2)An action space suitable for this research topic is designed,and the degree of freedom of the researchobjectisdetermined,sothattheagentcanmoveinthesceneanddisplaytheperspectivescreen.(3)A reward function that can effectively motivate the virtual camera to quickly realize autonomous control is designed,which improves the decision-making efficiency of the virtual camera,and is used to connecttheobjectdetectionmoduleandthereinforcementlearningmodule.(4)Three experimentalscenes were builtthrough the Unity3 Ddevelopment engine,and in thesethree virtual scenes,the Vanilla Policy Gradient algorithm(VPG)and the Trust Region Policy Optimization algorithm(TRPO)were used to realize the autonomous control of the virtual camera from the start to the target.Finally,the experimental results in each scene are compared and analyzed to verify the feasibility andperformanceofthetwoalgorithmsintheapplicationofthisthesis.This research project uses the object detection algorithm to identify the virtual exhibits in the virtual scene without obtaining the spatial position of the target and the virtual camera,provides the state support required for autonomous control of reinforcement learning,and realizes the autonomous control of the virtual camera through VPG and TRPO algorithms respectively,and displays the exhibits in the virtual scene.The experimental results of the two algorithms in three virtual scenarios show that the TRPO algorithm is more suitable for complex virtual scenarios,and it is more adaptable,higher in cumulative rewardvalueandbetterinperformancethanthe VPGalgorithminthisapplication.