RI: Small: Exploiting Correlated Sparsity Pattern Change in Dynamic Vision Problems

点击翻译按钮获取中文摘要

This project develops a new framework to solve a large class of dynamic vision problems by exploiting correlated sparsity pattern change in the appropriate domain. The focus is on high-dimensional visual tracking problems such as deformable contour tracking or target tracking in the presence of significant illumination changes. These are difficult because of the high dimensionality and because the observation models are highly nonlinear and/or non-Gaussian due to clutter, occlusions or low contrast. However, in most such problems, even though the state (e.g., contour deformation or illumination) is high-dimensional, at any given time, most change occurs in only a few principal directions. In a long sequence, this set of directions can gradually change over time. Most existing methods need a set of past state estimates to estimate this change on-the-fly while tracking noisy or nonlinear systems. The research team provides a completely new solution to this difficult problem by re-interpreting it as a problem of "recursively reconstructing sparse state sequences with slow time-varying sparsity patterns" and tapping into ideas from their ongoing recursive sparse recovery work.The research of this project enriches the knowledge base of computer vision and can be applied to many different applications such as medical image analysis and video surveillance. The project provides research opportunities for graduate students and involves undergraduate students, including under-represented minorities, through summer, senior design projects and REU projects.