面向形状理解的三维形状统计分析技术

With the rapid development of both the three-dimensional acquisition and processing technology and applications, massive three-dimensional content has brought new challenges to the storage, analysis, retrieval and synthesis of new content. In recent years, the work on three-dimensional analysis focuses on the calculation of the local features and global features. However, due to the gap between the low-level features and high-level semantics, shape understanding of theoretical and applied research is still relatively poor. Latest researches focus on a combination of statistical machine learning and three-dimensional content analysis and semantic understanding, so we will try to deepen this combination. The project study semantic understanding of the three-dimensional shape of the three key issues: shape complexity based on mutual information, three-dimensional segmentation and labeling based on the Conditional Random Field and the Active Shape Modeling methods, and the extraction of three-dimensional shape hierarchy guided by multi-level semantics. On the basis of these three key issues, we propose the project which shows the hierarchy of three-dimensional shape extraction framework based on statistical analysis of three-dimensional shape technology . Technological contributions of the proposal are: design and evaluation of shape complexity measurement, three-dimensional segmentation and labeling based on the Sparse Representation and the Active Shape Modeling, extraction of three-dimensional shape hierarchy guided by multi-level semantics, and learning and assembly technology of three-dimensional deformable shape of the discrete structure.

随着三维获取技术和三维内容应用的快速发展，海量三维内容给存储、分析、检索和新内容合成等相关研究带来新的挑战。近年来三维分析工作侧重于局部特征和全局特征的计算。然而，由于低层特征和高层语义之间的还有鸿沟，针对形状理解的理论和应用研究还比较少。最新研究热点是统计机器学习和三维内容分析和语义理解中的结合，我们希望加深这个结合。本项目研究三维形状语义理解的三个关键问题：基于互信息的形状复杂度计算、基于条件随机场和主动模型的三维分割与标注、以及多级语义引导的三维形状层次结构提取。在这三个关键问题研究的基础上，本项目提出了基于三维形状统计分析技术的三维形状层次结构提取框架的研究方案。项目的技术创新有四点：衡量形状复杂度的度量设计方法，三维形状稀疏表示和主动形状模型确定的形状分割与标注方法，基于分级语义的三维形状层次结构提取形状，三维形状离散结构变化的学习与组装技术。

三维数据分析在3D打印、3D游戏和电影、无人驾驶、智能监控等领域都有较好的应用前景。从技术层面看，三维模型的形状分析和三维场景理解是这些应用得以深入的基础研究。本项目按照任务书计划，在预期研究内容取得了多个方面的进展：提出了基于蓝噪声采样的随机点云的生成方法，为算法测试和方法评价提供了数据支撑；提出了基于互信息、信息熵和各向距离差异度的形状复杂度的度量方法；提出了概率混合策略、基于条件随机场与机器学习相结合的点云分割/分类与标注方法；通过基于圆柱形状检索的树枝结构提取方法以及基于分级粒子流技术的三维树木重建方法，并用于植物动态生长模拟与生长过程可视化。在三维形状复杂度度量、三维点云模型的形状特征计算及其在三维分割与重建中的应用、基于分级语义的三维树枝形状结构提取与生长参数估计、三维植物模型的离散结构变化与组装新植物模型并用于植物生长模拟方面有较好的突破与创新。依托本项目已经发表论文27篇，包括11篇国际期刊论文，6篇国内期刊论文和10篇国际会议论文。其中SCI检索论文10篇，获得会议论文最佳论文提名奖1篇。申请国家发明专利3个，其中已经授权1个。登记软件著作权4个。培养硕博士研究生8人，其中获评系统仿真学会的优秀博士论文1篇。项目组成员获省部级和市级学术奖励各1项。