An Skeletal Animation Subsystem Base On Neural Network Motion Synthesis

Skinned skeleton is often used for character animation of 3D simulation system.To get a natural and realistic movement of character animation,massive motion capture and manually data processing are needed.such inaccurate data like pose or speed of movement is inconveniently to compute during the creative process of motion synthesis.Some neural network like locomotion synthesis network is already able to produce a long sequence of periodic motion,Phase Functioned Neural Networks(PFNN)can predict a realistic human walking movement adapted to a rough terrain.Furthermore,Mode Adaptive Neural Networks(MANN)is capable to generate quadruped animation walking on various terrains.By porting the neural network of motion synthesis to a widely used 3D rendering engine with the system provided by this thesis,the efficiency of creative process can be improved.The research contents are as follows.First,the framework of the skeletal animation subsystem designed in this thesis with the neural network technology can porting to any general 3D engines.Supporting new neural networks extension and multiple instances make the subsystem easy to use and more flexibility.Second,by porting PFNN and MANN neural network model to the subsystem,including porting of resource and skinning technology,the character movements produced by the subsystem become more natural and various.Third,we build a verify scene for this subsystem using multiple instances of those neural network models and adding script control.The function and performance are analyzed after running of the scene.It is verified that the actual operation effect of this subsystem basically meets the requirements of prototype.This thesis proposed a solution for that most character skeletal movements of games and animations are still designed by artists manually in a traditional way for now because of no full-equipped toolchain of neural networks of motion synthesis.In addition,the system can be a powerful supporting tool for simulating the road environment of driverless cars with pedestrians and motion learning of robot.