Interactive Object Physics Simulation Solver In Virtual Scene

Digital humans are gradually playing an increasingly important role in digital life.When digital humans interact with objects in virtual scenes,especially in the field of real-time rendering,they often face mold penetration,unnatural movement,and weak interaction with objects,etc.Therefore,providing more efficient and realistic physical calculations has gradually become a hot spot in computer animation research in the field of real-time rendering in recent years,and the research on digital humans performing physical interaction calculations in virtual scenes is also important for technologies such as virtual reality and future human digital life is of great significance.The calculation objects of the physical simulation solver in the virtual scene are mainly divided into three categories:rigid body,cloth and elastic body.This thesis proposes a hybrid algorithm based on physics and neural network for the more difficult cloth and elastic body calculation problems.The cloth solver and the elastic body solver calculate the data compression optimization scheme of the subspace neural network,and carry out the simulation experiment according to the designed system model and algorithm.The main work of the thesis is as follows:1)The hybrid cloth solver proposes an unsupervised neural network cloth calculation algorithm based on pose space deformation and a physical cloth calculation algorithm based on proxy mesh from the perspective of machine learning and traditional mechanical calculation,and through weight mixing the two are organically combined.At present,the solution scheme based on machine learning is mainly concentrated in the simulation of human body folds,and it faces great difficulties in interacting with the surrounding environment,and there are still many distortions in the simulation effect of non-fitting clothes.The part of the neural network in the hybrid solver by introducing a loss function that perceives bone velocity changes and an external force loss function,the distortion problem solved by the current neural network is optimized and the simulation of external forces such as wind force is expanded;the current solution based on traditional mechanics mainly faces resolution.In order to balance the performance with the calculation speed,the mechanical calculation part of the hybrid solver uses proxy grid technology to down-sample and oversample the original cloth to balance performance and calculation accuracy.At the same time,the hybrid solver also realizes the weight mixing of the two at the shader level,so that the final solver can combine the advantages of the fast calculation of the machine learning solution and the ability of the mechanical solution to interact with the surrounding environment.From the results,the comparison the current scheme is more efficient and practical,and the scheme is innovative and original.2)The data compression optimization scheme of the subspace neural network in the elastic body calculation is mainly aimed at the problems of too many solution vertices,large model complexity,and slow front-end reasoning in the elastic body calculation.Model surface reduction algorithm and subspace neural network data compression algorithm.Since the interactive part of the interactive object in the virtual scene often only occupies a part of the object model and has a large relationship with the visible viewing angle,the surface reduction scheme mainly uses the calculation of the lower envelope to select the visible area of the model under the specified viewing angle.and reconstruction;at the same time,for the interactive objects of rigid body and near-rigid body in the scene,the data compression optimization scheme no longer uses the position and speed of the model vertices for rendering and training,and optimizes the design of subspace neural network simulation from the perspective of data structure;Algorithm models also have specific optimization methods when deploying front-end reasoning.The data compression optimization scheme proposes optimization schemes from the perspectives of neural network pruning,instruction set acceleration,and model accuracy selection,and performs simulations.From the results,it can be seen that the data compression optimization scheme of the subspace neural network can effectively reduce the amount of data in the training process of the neural network,improve computing efficiency and optimize the model training process and front-end reasoning deployment.