Computer animation is the process which uses several techniques of computer graphics to create movements of 2D and 3D objects. For 3D animation, animation authoring comprises of several steps including modeling, rigging, and animation. Typically, these processes need to be completed manually. Skilled animators are required to complete these tasks. Animation authoring might be infeasible for nonprofessional users. The main goal of this dissertation is to bring animation authoring to users with any skill level by developing a computer animation system that can be used by everyone. Such system should enable us to develop several interesting applications, for example, one may wish to develop a storytelling system that automatically generates animation after users tell a story or play with a puppet. Such application requires a novel computer animation system which can be operated with simple interactions by anyone, even nonprofessional users.

Nowadays, several systems have been proposed to reduce the difficulty of computer animation authoring. 3D scanners and several modeling systems have been developed to assist users in the modeling step. In the animation step, motion capture systems have been used in animation production to create the motions for 3D characters for more than a decade. Furthermore, there are several repositories for 3D models and motions, which users can obtain easily. However, it is not easy to combine 3D models and motions to generate animation. The 3D model must first be rigged to make it ready to be animated. The purpose of the rigging process is to identify the skeleton structure that can be used to deform the pose of the 3D model. Normally, this process must be performed manually, and it is too difficult for nonprofessional users. For the motion capture systems, although they can be used to generate realistic motion data for animation, such motion data can only be achieved with high-cost systems and several difficult post-processing tasks. With this reason, it may not be suitable for nonprofessional users to use such systems to obtain motions of a single action with wide variety of styles.

The problems stated above prevent nonprofessional users from performing the animation authoring task. Therefore, we propose a novel computer animation system that is suitable for nonprofessional users. The proposed system enables users to implement computer animation with a small amount of efforts, and allows users to create an animation with their chosen 3D model even if it has not been rigged. Furthermore, it provides the possibility for users to create a new motion with simple interactions. Given these capabilities, the system has high flexibility which allows users to create computer animations with any 3D model and motion. The proposed animation system consists of three modules: animation module, modeling module, and motion module. The animation module is the module for generating computer animation by re-targeting (transferring) motion data into a 3D model. We use the framework of an existing motion re-targeting based animation system to develop the animation module for our proposed system. In this module, 3D models and motions are retrieved from model and motion databases before performing the re-targeting process. Because the number of motions in the motion database is quite large, searching suitable motion manually is time-consuming, and not convenient for users. We propose to use a puppet interface to generate the query information for retrieving suitable motions from the database. The motion database usually stores several motion sequences, which each of them conveys only a single action, that is, the retrieved motions must be connected to generate a single motion sequence for animation. We propose a new motion concatenation using the cubic Bezier interpolation to create transitions for any consecutive motions. The modeling module is included in the system to automatically generate a rigged 3D model from the input 3D model (if it has not been rigged). We propose a novel template-based automatic rigging algorithm for this modeling module, so users can obtain rigged 3D model without human intervention. The motion module is provided to enable users to create a new motion by editing the example motions in the database. We propose a simple motion editing technique using timeline-based interface which should be suitable for this module.

For the automatic rigging process, we propose an algorithm that automatically embeds the suitable template skeleton into a 3D model. We develop a classification algorithm based on the characteristic of real animals to identify the character class of the input 3D model, and extract the anatomical semantics of each part of the model. The classification rules are applied to the curve skeleton extracted from the input 3D model to obtain the classification results. This classification rules are not restricted to only the standard pose, orientation, or topology, which means 3D models created from any system can be rigged by our proposed method. To reduce computation time, we propose the Skeleton Growing algorithm to extract a curve skeleton from the pseudo-normal vector field constructed inside the 3D model. This algorithm can generate the curve skeleton with short computation time without user intervention. Therefore, the proposed technique can generate a rigged 3D model automatically using a short amount of processing time.

The proposed motion editing techniques have been developed for nonprofessional users in the way that they can edit a previously stored motion to their desired motion with simple interactions. The method firstly extracts the key poses from the motion capture data. It then visualizes the key poses on the timeline. We decide to create the timeline in the 3D environment instead of the 2D environment used in existing works to eliminate the ambiguity problem caused by inappropriate viewing directions. We proposed interaction techniques for motion editing functions that employ several editing algorithms. The proposed editing system provides several editing functions that can be used to edit both the geometry (pose) and temporal information of the motion data. Each editing function is developed with consideration of reducing the difficulty and the completion time of the editing task, so nonprofessional users can use the proposed editing system to create a new motion data by editing example motions.

We propose a motion concatenation algorithm that employs parametric motion synthesis to generate motion data for animation. Two or more motions can be connected to generate a single long motion. In contrast to existing works which also use the technique of parametric motion synthesis, the proposed method can generate transition for any consecutive motions even though there is no intersection region of their parameter spaces. We use the cubic Bezier interpolation to generate the transition between consecutive motions, so the transition is quite smooth even if the poses between two motions are quite different. We propose an algorithm to pre-compute all possible poses used to control the interpolation. The process to retrieve the suitable poses for interpolation, therefore, can be completed with low computation complexity. Furthermore, we also propose an algorithm to create feature vector for motion clustering. Because we employ the parametric motion synthesis, the example motions must be clustered into motion groups such that each motion group contains motions with similar actions. We propose the feature vector, which is calculated from kinetic energy of each joint of the motion data for this clustering process. The proposed method is faster than the typical methods that calculate the difference between motions frame-by-frame. Moreover, this method is simple because the clustering result can be obtained without any query motion from the users. The process of concatenating motions can be completed without any user intervention, so it is suitable for the proposed animation system for nonprofessional users.

In this dissertation, we also investigate and discuss about the suitable user interface for the computer animation system. We propose to use a puppet interface to generate computer animation. Playing with the puppet is a simple interaction which users should be able to perform with a low learning curve. Such interface can be used to retrieve all required motions from the motion database, which will be concatenated into a single motion before being re-targeted into a rigged 3D model for generating computer animation. Users can use the puppet to retrieve motion data by posing the puppet with the poses of their required motion. The pose information of the puppet is composed of rotation angles of all the joints of the puppet. To obtain such pose information, we attach several visible markers on the puppet. The marker tracking algorithm can identify the orientation of each marker, which can be used to calculate the rotation angle of its associated joint. The marker occlusion problem is also addressed in this dissertation. The proposed algorithm can estimate the rotation angle of each joint by determining the position and orientation of the adjacent markers of the occluded marker.

The contributions of this dissertation are the four algorithms explained above. With the system configuration and all the proposed algorithms, the computer animation system for nonprofessional users can be developed. The proposed system allows users to create animation automatically using the desired 3D models and motions. The experimental results demonstrate the performance of each of the proposed algorithms in both quantitative and qualitative areas. The results confirm that the proposed algorithms provide reliable outcomes for the proposed computer animation system, and all interactions required by each of them are simple enough that nonprofessional users can perform. We believe that with this proposal of the computer animation system, the animation authoring can be brought to all levels of users. Several interesting applications related to computer animation can be developed based on the proposed computer animation system. Such applications should contribute many benefits to both academic and industrial sectors.

本論文は、「A Novel Computer Animation System for Nonprofessional Users 」(非専門家ユーザのための新しいコンピュータアニメーションシステム)と題し、英文で書かれており、10章よりなる。近年、コンピュータアニメーション技術に関する関心が高まっている。しかし、アニメーション作成ソフトウェアを用いた作業は決して容易ではない。具体的には、対象となる3次元モデルからのスケルトン抽出や関節位置の決定等のリギング、スケルトンの動きに応じてモデル表面を違和感なく変化させるスキニング、モデルの各部位やモデル全体に対し自然でスムーズな動きを与えるモーション生成などの作業に多大な時間を要する。よって、コンピュータアニメーション作成は専門家にとっても大変骨の折れる作業である。本論文では、このような現状を克服する新たなフレームワークを提案し、個々の作業を自動化またはユーザの負荷軽減をするシステムの実装と評価を行っている。

第1章は、「Introduction(序論)」と題し、本論文の背景、動機、貢献を明らかにするとともに、本論文の構成について述べている。

第2章は、「Related Work (関連研究)」と題し、コンピュータアニメーション作成を支援するシステムや要素技術について、具体的な例を挙げつつ説明している。

第3章は、「System Overview (提案システムの概要)」と題し、本論文で構築されるシステムのフレームワークについて議論し、提案システムを構成するモデリングモジュール、モーションモジュール、アニメーションモジュールの機能について述べている。

第4章は、「A Novel Template-based Automatic Rigging Algorithm (テンプレートに基づく新しい自動リギングアルゴリズム)」と題し、擬似法線ベクトル法による3Dモデルからの高速かつ正確なスケルトン抽出、スケルトン分類、テンプレートを用いた関節位置決定を自動的に行うリギングアルゴリズムについて述べている。

第5章は、「Motion-editing Technique using a Timeline-based Interface (タイムラインインタフェースを用いたモーション編集手法)」と題し、モーション編集作業を軽減するための時空間伝播手法を提案し、3Dモデルを直接操作できるユーザインタフェースについて述べている。

第6章は、「Parametric-motion Concatenation using Bezier Interpolation(ベジェ補間を用いたパラメトリックモーション連結)」と題し、ユーザの要求に対応して複数モーションを連結する手法について述べている。各関節の運動エネルギーを特徴量とするモーションデータのクラスタリングとベジェ補間を組み合わせることで、従来手法より高速なモーション連結を行えることを示している。

第7章は、「A Puppet Interface for a Novel Computer Animation System for Nonprofessional Users (非専門家ユーザのための新しいコンピュータアニメーションシステム向け人形型インタフェース)」は、より容易にモーション編集やモーション検索を行える人形型の物理インタフェースを導入し、視覚マーカのオクルージョンに対するロバスト性を高めるアルゴリズムについて述べている。

第8章は「Experimental Results (実験結果)」と題し、第4章から7章までで提案した個々の手法に対する実験結果を基に、その効果について議論している。

第9章は「Discussion (考察)」は、前章の実験結果を基に各提案手法の利点と限界を明らかにするとともに、それらの改善点等について述べている。

第10章「Conclusion (結論)」は、本論文の結論である。

以上を要するに、本論文は、非専門家ユーザが容易にコンピュータアニメーションを作成できるシステムおよびフレームワークを提案、実装、評価し、その有用性を示したものであり、情報工学の発展に貢献するところが少なくない。

よって本論文は博士(工学)の学位請求論文として合格と認められる。