Implementation of Efficient Computer Vision Based on a Robust Subspace Method

从计算机视觉的角度研究了子空间方法的一种新框架。在这个框架中，特征空间由归一化特征空间的齐次表示来表示，其中归一化意味着11-归一化。然后，通过考虑所有有效像素，通过将图像线性投影到齐次特征空间，有效地实现了给定图像到给定归一化特征空间的部分投影。当孤立点检测模块与简单部分投影相结合时，这一简单的公式可以扩展到稳健的部分投影，有效的部分投影可以促进稳健的人脸识别、实时的人脸/目标跟踪等应用。在人脸识别中，引入了并行部分投影的概念，实现了自然环境下人脸识别的稳健机制。在实时跟踪中，引入了稀疏模板匹配的概念，实现了模板与输入图像的有效匹配。将稀疏模板匹配应用于粒子滤波框架中，实现了一种非常高效的跟踪器。在已知特征模板的情况下，稀疏模板匹配也可以推广到稀疏特征模板匹配，特征模板匹配也可以应用到粒子滤波中。研究中还对其他一些扩展进行了研究。

A novel framework of the subspace method is investigated from a viewpoint of computer vision. In this framework, an eigenspace is formulated by a homogeneous representation of a normalized eigenspace., where the normalization means 11-normalization. A partial projection of a given image onto a given normalized eigenspace is then efficiently made by a linear projection of the image onto a homogeneous eigenspace by taking all the effective pixels into account. This simple formulation can be extended to robust partial projection when an outlier detection module is combined with the simple partial projection.The efficient partial projection can facilitate robust face recognition, real-time face/object tracking, and other applications. In the face recognition, a concept of parallel partial projections is introduced for implementation of robust mechanism for the face recognition in natural environments. In the real-time tracking, a concept of sparse template matching is introduced for accomplishing an efficient matching of a template and an input image. When the sparse template matching is utilized in a framework of particle filter, a very efficient tracker is implemented. The sparse template matching can also be generalized to sparse eigen-template matching when the eigen-template is given in prior, and the eigentemplate matching can also be utilized in the particle filter. Some other extensions are also investigated in the research.