From Optical Flow to Image Deformations

本研究旨在证明视觉运动分析可以通过(1)用图像变形（它们彼此之间的运动）取代光流分析（图像点的运动）和(2)跨越更大的视场来改进。图像变形的优点类似于差分放大器在噪声背景下检测微弱信号的优点：如果噪声对两个输入的影响相等，则通过取它们的差来抵消。在类比中，摄像机旋转引起的像移是噪声，平移引起的像移是有用信号。关于视场，文献表明，60度或更大的视角几乎会导致重构问题条件的量子改善。这一线索是由相机运动引起的广角图像变形的形状重建的理论和实证研究所追求的。图像变形和宽视场这两个概念已经在初步研究中取得了成功，目前正在进行理论和实验研究。最后的贡献将是一个多帧形状和运动重建方法，可靠地工作，并为透视图像。

This research aims at demonstrating that visual motion analysis can be improved (1) by replacing the analysis of optical flow (the motion of image points) with that of image deformations (their motion relative to each other) and (2) by spanning a wider field of view. The advantages of image deformations are akin to those of differential amplifiers for the detection of weak signals in a noisy background: if noise affects both inputs equally, it is canceled by taking their difference. In the analogy, image motion caused by camera rotation is the noise, and that caused by translation is the useful signal. As to the field of view, the literature shows that a viewing angle of 60 degrees or more leads almost to a quantum improvement in the conditioning of the reconstruction problem. This lead is being pursued by a theoretical and empirical investigation of shape reconstruction from the deformations of wide-angle images induced by camera motion. The two ideas of image deformations and a wide field of view, already proven successful in preliminary studies, are being investigated theoretically and experimentally. The final contribution will be a multiframe shape and motion reconstruction method that works reliably and for perspective images.