基于距离图像局部特征的三维形变目标识别技术

3D object recognition in range images has became a popular research topic in the area of computer vision, it has a number of applications in both civilian and military areas including remote sensing, automatic navigation, precision guidance, and battlefield surveillance. Different from existing algorithms which focused on the recognition of rigid objects in small datasets, the proposed research intends to investigate several key technologies for deformable object recognition in incremental data with complex scenes. A range image is regarded as a 2D manifold embedded in the 3D Euclidean space, and the geometric diffusion technique is used to construct an intrinsic scale space for the range image, deformation-invariant adaptive-scale keypoints are then extracted from the intrinsic scale space. The geometric, color, and texture information is studied to extract efficient and robust feature descriptors for the neighborhoods of keypoints, and the subspace learning theory is employed to perform feature fusion of multimodal information and also to perform dimension reduction. An incremental Self-Organizing Map Tree (SOMT) is investigated to perform efficient feature matching between a scene and models, the mixture probability statistical model and the large deformation diffeomorphic non-rigid transformation model are developed to achieve fast and robust fine registration between the scene and a model. The proposed research can perform detection and recognition of 3D rigid, articulated and deformable objects in incremental and incomplete data.

针对距离图像的三维目标识别技术已成为计算机视觉领域新的研究热点，其在遥感测量、自动导航、精确制导和战场侦察等军民用领域有着广泛的应用前景。与已有三维目标识别算法主要解决小规模数据集下的刚性目标识别不同，本项目旨在研究数据增量条件下复杂场景环境中针对距离图像的三维形变目标识别若干关键技术。拟将距离图像视为嵌入在三维欧氏空间中的二维黎曼流形，研究几何扩散方法实现内蕴尺度空间构建，提取对形变稳健的自适应尺度关键点；研究基于形状、颜色及纹理信息的关键点邻域高效稳健特征描述算法，并采用子空间学习理论实现多模态信息的特征融合和降维；研究增量自组织映射神经网络树技术实现快速高效的场景与模型特征匹配，开发基于混合概率统计模型和大形变微分同胚变换模型的点模式匹配算法完成快速鲁棒的场景与模型精确匹配，解决增量数据和不完备数据条件下的三维刚性目标、关节目标以及等距形变目标检测和识别问题。

针对距离图像的三维目标识别技术已成为计算机视觉领域新的研究热点，其在遥感测量、自动导航、精确制导和战场侦察等军民用领域有着广泛的应用前景。与已有三维.目标识别算法主要解决小规模数据集下的刚性目标识别不同，本项目旨在研究复杂场景环境中针对距离图像的三维形变目标识别若干关键技术。在点云局部特征描述方面，首先提出了一种全新的点云局部特征描述算法框架，并依据该算法框架设计了自旋图三元组（TriSI）点云局部特征描述子，其在鉴别力和稳健性方面均获得了优异的性能。此外，针对RGB-D数据，提出了一种融合颜色和几何信息的局部二值特征描述子。在仿射变换下的点模式匹配方法研究中，为了解决传统的一致性点漂移算法存在的局部最优性和收敛速度随点集大小增加而下降等问题，提出了一种新的基于全局最优的快速一致性点漂移算法。在非刚体变换下的有标记点模式匹配方法研究中，当归纳分析了经典的非刚体几何变换模型以及经典的基于微分同胚的非刚体变换模型所存在的问题后，提出了一种新的基于恒定动量矢量的快速大形变微分同胚非刚体变换有标记点模式匹配算法。在三维目标识别方面，首先针对遮挡和背景干扰下的目标识别问题，提出了基于模型库的层次化三维目标识别算法。针对多视图融合三维目标识别中信息冗余的问题，提出了一种融合视图显著性与多视图表征的卷积神经网络，并用于提取三维目标的有效特征。