Study On Synthesis Of Character Animation And Its Stylization

Charcter animation is an important part of computer graphics community. The motions and behaviors of virtual humans are not only required to be realistic, but also expected to be more expressive and life-like. In order to meet this requirement, this thesis focuses on character animation synthesis and stylization, including the methods of motion style editing based on style subspaces, data-driven stylistic human motion synthesis combining conventional inverse kinematics, motion generation based on manifold learning. Moreover, we also develop an expressive character animation creation platform, which endues virtual humans with personality and emotion. Our study provides animators with a set of tools for fast and convenient creation of expressive character animation.On stylistic human motion synthesis, this thesis presents a low-dimensional subspace-based motion style editing method. An unsupervised machine learning technique, Independent Feature Subspace Analysis, which is a combination of multi-dimensional independent component analysis and invariant feature subspace, is used. This technique can divide motion data into several mutually independent subspaces and one of these subspaces is defined as style subspace. The animator can tune, transfer and merge the style subspaces to synthesize new animation clips with various styles. The main advantage of this method is only one motion pair is required to train the motion model. This method is intuitive and easy to use. The experimental results demonstrate that the method is effective for cyclic motions, e.g. locomotions.In order to synthesize stylistic motions that meet specified constraints, we improve our previous low-dimensional motion model in motion vector space. Based on this new motion model, an optimazation-based method for synthesing sytlistic human motions satisfying user-specific geometric constraints is proposed in this thesis. This method relates full-body Jacobian matrix to low-dimensional Jacobian matrix using chain rule, so that inverse kinematics problem can be resolved in the low-dimensional space. Additionally, the style subspace can respond to the input sytle parameter. Therefore, this method is able to preserve the motion style while resolve inverse kinematics problems. Boxing motions are taken as examples in our experiments. The experimental results prove that this method is efficient and can be used for interactive animation editing and real-time games.However, it is sometimes hard to obtain satisfactory results by using linear subspace technique when constraints increasing. Therefore, a technique called Principal Geodestic Analysis that combines riemann manifold and lie group is adopted to train low-dimensional motion model. This method firstly projects motion data on geodesics on a Riemann manifold to find the non-linear embedding under the data, then synthesize new motions satisfying constraints through optimazation. Compared to linear subspace based technique, principal geodestic analysis has much stronger generalization capacity and can depict motion data structure more essentially. Compared to other non-linear dimensionality reduction technique, it has more concise analytical form suitable for optimazation computation.Finally, an expressive character animation platform, which endures virtual human with personality and emotion is developed. There are two main levels in this platform. One is a subspace based human motion engine at low level, another is a behavioral animation authoring sytstem at high level. In this platform, stylistic human motion sythesis techniques, behavioral modeling and affective computing techniques are integrated together by a map between high-level sementics and low-level raw motion data. This platform provides animators with a set of tools for fast and convenient expressive character animation creation.