Image motion in computer vision: Variational processing and experimental verification

本研究的目的是在变分范式下为计算机视觉中图像运动处理的基础学科提供一个通用的算法框架。主题是运动估计、检测、跟踪和三维解释。运动估计是根据图像的时空变化来评估光速度场。运动检测确定图像中被运动物体覆盖的区域，即将图像分割为运动物体的前景及其补体。跟踪是通过图像序列跟踪运动物体的过程，三维解释恢复产生图像运动的物体在空间中的形状和运动。变分方法从优化的角度来表述问题，即，作为目标函数的优化，它表达了表征期望解决方案的所有约束。因此，他们可以计算出非常精确的解决方案，而且效率很高。尽管变分方法具有明显的优势，但由于传统的启发式处理方案已经解决了许多重要问题，因此在运动分析中还没有系统地研究变分方法。这项研究将探讨目前用非变分方法处理的问题，但也将解决重要的开放问题。研究将包括对每个问题的变分分析，其中包括目标函数的设计和对其优化的有效方法的询问，例如参数图切割和通过水平集的变分演算。所得到的公式将通过使用来自自由访问的数据库的图像序列的详细比较实验来验证，这些数据库专门用于验证图像运动处理算法。这项研究的预期结果是一个统一的高效，易于使用，并记录良好的运动分析算法。

The purpose of this research is to provide a common algorithmic framework under the variational paradigm of the fundamental subjects of image motion processing in computer vision. The subjects are motion estimation, detection, tracking, and three-dimensional interpretation. Motion estimation evaluates the optical velocity field from the image spatiotemporal variations. Motion detection determines the region in an image covered by the moving objects, i.e., it segments the image into a foreground of moving objects and its complement. Tracking is the process of following moving objects through an image sequence, and three-dimensional (3D) interpretation recovers the shape and motion of the objects in space which produced the image motion. Variational methods formulate problems from an optimization viewpoint, i.e., as the optimization of an objective functional which expresses all of the constraints that characterize the desired solution. As a result, they can compute very accurate solutions and do so efficiently. In spite of their manifest advantage, variational methods have not been systematically investigated in motion analysis as there are currently many important problems that have been addressed by traditional heuristic processing schemes. This research will investigate problems currently treated by non-variational methods but will also address important open problems. The investigations will consist of a variational analysis of each problem, which includes the design of the objective functional and an inquiry about efficient means of its optimization, such as parametric graph cuts and the calculus of variations via level sets. The resulting formulations will be verified by detailed comparative experimentation using image sequences from freely accessible databases dedicated to the validation of image motion processing algorithms. The anticipated outcome of this research is a unified body of efficient, easily usable, and well documented algorithms for motion analysis.