Research On Evolutionary Strategy For Human Physics-Based Motion Generation

Motion generation of animated characters is one of the hot issues in the field of computer graphics.With the booming development of multimedia games and movie special effects industries,people's requirements for the reality of character animation motion continually increase.However,it is still a challenging research problem to generate high-quality,natural and realistic human motion.In recent years,character animation based on physics simulation has received extensive attention from researchers since it can generate real motion satisfying physical laws and respond to environmental changes.However,the physical model of human body is nonlinear,high-dimensional and strong joint coupling,which makes the physical motion of human body difficult to solve;On the other hand,most of the generation methods refer to related technologies in the field of robots,which results in the mechanical stiffness of generated motion.Thus it is hard to guarantee the fidelity.In view of the above problems,this paper carries out the research of physics-based human motion generation,including:firstly using motion capture data as the reference sequence for optimizing controller to ensure the naturalness and fidelity of generated motion;secondly constructing the human physical motion generation model for constrained solving and proposing the feedback controller based on linear compensation to reduce the difficulty of optimization;finally using the optimization algorithm based on evolutionary strategy to optimize the open-loop controller and feedback controller to generate natural and realistic human physical motion with feedback response.The main work is summarized as follows:(1)A human physical motion generation model for constraint solving is proposed.Firstly,the human body is physically modeled,and then a feedback controller based on linear compensation is proposed to reduce the difficulty of solving the feedback controller,which is based on the open-loop controller;Secondly,according to the characteristics of human motion and physical model,a set of objective functions for measuring motion effects is designed,and the constraint conditions that need to be satisfied are summarized;Finally,aiming at the problem of redundant collision detection in the generative model,a hybrid bounding box collision detection algorithm based on union-find sets is proposed to improve the physical simulation speed.(2)An open-loop controller optimizing algorithm based on clustering selection is proposed.On the basis of the model of human physical motion generation,the open-loop controller for a variety of motions is solved by the optimization algorithm combining clustering selection and intelligent evolution strategies.The sub-space partitioning clustering algorithm selects the optimal individual in the respective subspaces while ensuring the difference between the samples,thereby reduces the number of candidate solutions and improves the convergence speed of the algorithm.On this basis,the sliding window is utilized to solve multiple target poses uniformly,which reduces the stiffness of physical motion and speeds up the optimization.Compared with the existing algorithms,the proposed algorithm not only makes the human physical model better track the motion data,but also improves the robustness,time performance and stability.(3)A multi-objective algorithm for feedback motion controller is proposed.Based on the open-loop controller optimization,a multi-objective solving algorithm for feedback motion controller is designed and developed.Using the idea of multi-objective evolutionary algorithm,constraint solving for multiple targets is conducted simultaneously,which eliminates a lot of parameter adjustment process.In addition,the multi-layer optimization algorithm based on regional density is proposed to uniformly select the offspring.The RBF-SVM classifier of forbidden area is constructed by using successful individuals which satisfy constraints and failure individuals which do not satisfy constraints,and the failure individuals are pre-filtered in the offspring,therefore the hit ratio of the success individual is improved.On this basis,a multi-stage physics optimization algorithm based on pruning is proposed.The unstable controllers are discarded by the optimizing window to speed up the physical optimization.Experiments show that the algorithm is better than the previous algorithms in terms of the indexes of hit ratio,IGD and objective function values,and our algorithm not only completes the optimization to the motion controllers,but also greatly improves the convergence speed.(4)A system of human physical-based motion generation based on evolutionary strategy is designed and developed.The system is developed on the previous proposed algorithms,and decomposes the controller optimization into open-loop controller and feedback controller.Compared with the direct optimization for feedback controller,therefore the difficulty of optimization is reduced.The controller not only has the ability of the open-loop controller to track motion capture data,but also has the ability of the feedback controller to respond to environmental changes.Therefore,the controller ensures the fidelity and interaction of the generated physical motion.Finally,the stylized physical motion under different skeletons is realized,which increases the diversity of physical movement.