OPTIC FLOW
OPTIC FLOW
复制标题
DOI：
10.1016/0042-6989(86)90078-7
复制
发表时间：
1986-01-01
期刊：
影响因子：
1.8
通讯作者：
KOENDERINK, JJ
中科院分区：
文献类型：
--
作者：
KOENDERINK, JJ
本文从概念和计算方面对“光流”主题进行了快速回顾。该理论根据空间感觉运动行为和感知的神经生理学和实验心理学的可能应用进行了评估。哪种检测器适合提取光流的各个方面的问题受到了特别关注。结果表明，可能性实际上比当前（甚至是坦率的推测）文献中反映的要多样化得多。有人认为，对图像细节的方向差异的相对时间变化敏感的系统特别适合于关于诸如包含在光流中的物体的三维形状的信息来分析光流。因此，已知大量存在于许多脊椎动物的初级视觉皮层中的方向敏感元件因此被认为是提取环境物体的固体形状的很可能的基质。我们认为，应该以与通常解释相同的热情来研究这种可能性，通常的解释认为该系统负责平面（即图像中定义的）形状轮廓的初始提取。为“运动结构问题”提供了一种新的部分解决方案，它不仅涵盖了物体存在刚性运动时形状提取的常见情况，而且还涵盖了更广泛的（非刚性）弯曲变形（例如发生在不可扩展壳体的非刚性变形中）。这些解决方案违反了众所周知的“运动结构定理”所需的所有条件：这些解决方案对于没有四元组点作为刚性结构移动的点配置以及仅来自两个视图的输入数据是可能的。数值示例说明了如何使用该算法从非常有限的输入数据预测对象的侧视图。
This paper offers a quick review of the subject of “optic flow” in its conceptual and computational aspects. The theory is evaluated in terms of possible applications in the neurophysiology and experimental psychology of spatial sensorymotor behaviour and perception. The problem of which kind of detector is suited to extract various aspects of optic flow is given special attention. It is shown that the possibilities are actually much more various than is reflected in the current (even the frankly speculative) literature. It is argued that a system that is sensitive to the relative time changes of the orientation differences of image details is especially suited for an analysis of the optic flow with regard to the information concerning the three dimensional shape of objects such as is contained in the flow. Thus the orientation sensitive elements that are known to be abundantly present in the primary visual cortex of many vertebrates are hereby implicated as a quite likely substrate for the extraction of the solid shape of environmental objects. In our opinion this possibility should be investigated with the same ardour as the usual interpretation, which holds this system responsible for the initial extraction of the contours of flat (i.e. defined in the image) shapes. A new, partial solution to the “structure from motion problem” is offered, that not only covers the usual case of shape extraction in the presence of rigid motions of the object, but also the much wider class of (non-rigid) bending deformations (such as occur in the non-rigid deformations of inextensible shells). These solutions violate all conditions required by the well known “structure from motion theorem”: the solutions are possible for point configurations in which no fourtuple of points moves as a rigid structure and for input data from merely two views. A numerical example illustrates how this algorithm can be used to predict side views of an object from very limited input data.