【摘要】 动态几何模型表示了随时间演变的动态对象,通过非刚性注册匹配离散的帧模型是重建动态几何模型的核心问题.文中提出了一种鲁棒的成对非刚性注册算法,算法分为显式对应关系计算与全局变形优化两步:第1步,分析源模型和目标模型的滑动特征,提取显著特征点,建立特征点间的对应关系;第2步,利用显示的对应关系,求解全局变形优化完成非刚性注册,通过最小化能量函数,实现源模型和目标模型的最优匹配.实验表明,对于扫描获取和人工合成的数据,新算法突破了小尺度变形的限制,可以完成大尺度变形模型的成对注册.更多还原

【Abstract】 For the discrete frames,non-rigid registration is the key problem to reconstruct one dynamic geometrical model,representing a time-varying dynamic object.The paper presents a robust pairwise non-rigid registration algorithm,which is mainly performed in two steps: explicit correspondence computation and global deformation optimization.Firstly,after an efficient detection and analysis of slippage features on the source and target models,the salient feature points are extracted unsupervised.Then the correspondences among the feature points are correctly established.Secondly,by using the explicit correspondences,the approach finishes the non-rigid registration by a single global deformation optimization.The optimization maximizes the region of overlap and the spatial coherence of the deformation while minimizing the optimized energy function.The algorithm is fully tested on a series of real data sets,which are available from two ways: captured by the scanner devices or manually created from synthetic deformation tools.The results demonstrate that the approach is highly versatile,and applicable to large-scale deformable scenes that could not be handled by alternative techniques assuming that inner-frame motions are small.更多还原