【作者】 王瑢瑢；

【导师】 王兆其；

【作者基本信息】 中国科学院研究生院（计算技术研究所） ， 计算机应用技术， 2006， 硕士

【摘要】 如何生成逼真的三维角色动画一直是计算机图形学、虚拟现实等领域的研究热点与难点。运动捕获作为三维角色动画生成的一项主流技术,具有数据逼真度高细节丰富等特点,目前已成为计算机游戏、电影特效、训练仿真系统,医学辅助分析等诸多领域的标准方法。然而该方法从原理上是三维运动的简单录制和重放,无法提供对运动数据更多的交互控制。这使得该方法的数据可重用性低,难以适应交互性强的应用。本文针对利用已有的运动捕获数据库可控地生成新运动的问题,即基于运动库的运动生成进行了深入研究。主要成果如下:第一,提出并实现了基于三维运动库和一阶概率转移模型的运动控制方法如何描述运动数据内在的时空结构,是基于运动库运动生成方法的关键所在,决定了运动数据能否简洁有效的根据外界约束生成新运动。本文提出并实现了基于三维运动库和一阶概率转移模型的运动控制方法。该方法把运动数据看作一阶马尔可夫过程,即当前运动状态只和前一运动状态有关。在运动库预处理中,通过度量运动片段间的姿态速度相似性,估计运动片段间转移概率,从而建立了运动库一阶概率转移模型。在检索运动库时,我们选择和前一片段转移概率最大的片段作为当前片段,以此类推,最终得到概率最大的运动序列。我们以广播体操为例,结合相应的视频为交互界面验证了本文方法的有效性。实验表明,该方法能够有效地反映运动数据间的时空关系,提高运动库检索的准确性。第二,研究并实现了基于多尺度框架的运动片段平滑连接算法,保证了运动生成效果的逼真自然。从运动库检索出来的运动片段通常并不连续,需要进行片段平滑连接才能保证最终的运动生成效果。数据连接方法很多,一般的插值方法即可胜任。然而,由于运动捕获数据不同于基于关键帧或物理仿真方法生成的运动数据,其丰富的运动细节(即高频分量)使得简单的插值方法难以稳定估计运动片段在连接处的速度,最终难以得到平滑连接的效果。如何既保证片段间的平滑连接又保证原有的运动细节不被破坏,是运动捕获数据连接的难点。本文采用多尺度运动数据分析方法,首先将运动数据分解成一个低频基运动信号和不同层次的高频细节分量,然后分别连接基运动信号和各层细节分量,最后将连接后细节分量逐层添加到新的基信号上,最终恢复得到连接后的运动。大量实验表明,通过对基信号和逐层细节分量的分别处理,平滑连接结果在连接处达到了c1连续性并忠实地保持了原有的运动细节。更多还原

【Abstract】 The generation of realistic 3D character animation is a popular and difficult issue in theareas of computer graphics and virtual reality. Motion capture, as an important technology in3D motion generation, can obtain highly-fidelity 3D recordings of the motion of a liveperformer, and thus has become the standard motion generation solution in applications likecomputer game, movie special effects, sports simulation systems, and visualization aids formedical analysis. However, motion capture does not offer an animator free interactive control,but only allow one to play back what has been recorded. This drawback makes the existingmotion data difficult to be reused, especially in interactive applications.This thesis focuses on the problem of how to generate new motion with existing motioncapture database under users’ guidance. The contributions of this thesis are as follows:1. A motion control method based on 3D motion database and first-order probabilitytransition model is proposed and implemented.How to describe the spatio-temporal structure underlying the motion database is crucial tothe problem of motion generation based on motion database. In this thesis, we model motion asa first-order Markov process, that is, the current motion status is totally dependant on theprevious status. In the preprocess phase of motion database, we first measure the similaritybetween motion clips, according to which we then estimate the transition probabilities betweenthem, and finally build the first-order probability transition model of motion database. In theon-line motion control phase, we search the preprocessed motion database, and choose the clipthat has the largest transition probability with the previous one to be the current clip. Byrepeating the search-and-choose process, a most probable motion sequence is found out. Weverify the effectiveness of our method by adopting sport exercise video and a video interface.The result shows that our method can describe the spatio-temporal structure underlying themotion data effectively and promote the correctness in motion database searching.2. To guarantee the reality of generated motion, a motion stitching method based onmultiresolution analysis framework is studied and implemented.Generally speaking, motion clips retrieved from motion database are not continuous witheach other, so a motion stitching method needs to be adopted to guarantee the smoothness andreality of generated motion sequences. There are many interpolation methods could be used toconcatenate motion clips. However, obtaining a robust estimate of velocity from motioncaptured data is difficult because these data usually oscillate to include fine details that maydistinguish the motion capture data from those data generated by keyframing and physicssimulation. The difficulty of motion capture data concatenation lies in two issues: how toguarantee seamless concatenation between clips and how to maintain the fine details of originalmotion. A seamless motion stitch method based on multiresolution analysis framework isadopted in this thesis. We first decompose motion data into a base signal and levels of detailsignal, then interpolate base signal and detail signals respectively, and finally compose thestitched base signal and detail signals into a highly detailed motion signal, that is the stitchedmotion sequence. Experiment shows that the stitched motion sequence has c1 continuousnessand maintains the original motion details faithfully.3. A motion generation platform based on motion database with a video interface isimplemented.Due to the success of motion capture method, highly-fidelity motion data has rapidlybecome popular and commercial available. If we could utilize existing motion capture databaseand generate new motion under the users’ guidance, we will be able to reduce the cost ofanimation production greatly and provide a brand new technology solution for interactivevideo game. In this thesis, we implement a motion generation platform based on motiondatabase with a video interface. In the platform, we first reconstruct the 3D poses from videokey frames under the user’s guidance and calculate joints angle curves, then search thedatabase to find out the most probable motion sequence that matches the reconstructed jointsangle curves, and finally seamless stitch the motion clips with multiresolution analysis method.We use sport exercise video and relevant motion database to test the effectiveness of theplatform. Experiment shows that the generated motion is visually compatible with videocontents.更多还原