Motion Recognition by Parts Using Affine-Invariant Spectral Signatures

该研究的总体目标是开发一种新的方法来识别运动物体或部分，并提取和分类它们的轨迹。本研究的一个前景应用是对人体肢体、手指等身体部位的手势和运动的识别。该方法基于图像块的图标识别，这些图像块对应于大致平面的物体表面。利用仿射不变谱特征（AISS）和频域多维索引，对每个表面的三维姿态进行独立识别。AISS是基于二维调制的新型高斯核，这些核被组织在不同方向的子集中。采用两级层次索引方案对柔性目标进行位姿不变识别。索引方案提供每个识别对象/部件的位姿信息。因此，可以直接从处理过的图像序列中获得全3D运动（包括旋转）。提出了一种利用物体/零件三维轨迹的多维索引进行运动识别的新方法。将人体部位的三维运动轨迹分割成运动向量，然后利用这些运动向量的有序子集构造一组序列指标。然后使用这些有序子集来投票选择模型

The general objective of the research is to develop a new approach for recognition of moving objects or parts as well as extracting and classifying their trajectories. A prospective application of this research is recognition of human gestures and motion of body parts, such as limbs, fingers, etc. The approach is based on iconic recognition of image patches that correspond to object surfaces that are roughly planar. Each surface is recognized separately invariant to its 3D pose, employing affine invariant spectral signatures (AISS) and multi-dimensional indexing in the frequency domain. The AISS are based on novel Gaussian kernels with 2-dimensional modulation which are organized in subsets of distinct orientations. A 2-level hierarchical indexing scheme is used for pose-invariant recognition of flexible objects. The indexing scheme provides pose information of each recognized object/part. Thus, full 3D motion (including rotations) is directly available from processed image sequences. A novel method is planned for motion recognition using multi-dimensional indexing of the 3D trajectories of objects/parts. The 3D trajectories of body parts are segmented into motion vectors and then a set of sequential indices is constructed using ordered subsets of these motion vectors. These ordered subsets are then used to vote for model