Behavioral Visual Motion Analysis

Yiannis Aloimonos, Zoran Duric, Cornelia Fermuller, Liuqing Huang, Ehud Rivlin, and Rajeev Sharma.
Behavioral Visual Motion Analysis.
In IUW, 521-541, 1992

Online Version

A pdf version is available for download.

Abstract

We propose here a new approach to addressing problems related to visual motion, namely the purposive approach. Instead of considering the various visual motion tasks as applications of the general structure from motion module, we consider them as independent problems and we directly seek solutions for them. As a result we can achieve unique and robust solutions without having to compute optic flow and without requiring a full reconstruction of the visual space, because it is not needed for the tasks. In the course of the exposition, we present novel solutions to various important visual tasks related to motion, such as the problems of motion detection by a moving observer, passive navigation, relative-depth computation, 3-D motion estimation, and visual interception, using as input only the spatial and temporal derivatives of the image intensity function. It turns out that the spatiotemporal derivatives of the image (i.e. the so-called normal flow) do not seem to be capable of solving the general "structure from motion" problem. They are, however, sufficient to provide robust algorithms for the solution of many interesting visual tasks that do not require the full solution, but only part of it. The ability to create robust nontrivial behaviors suggests the possibility that visual perception could be studied as intelligent behavior. We point out some of the benefits and drawbacks of this paradigm that studies vision as a set of behaviors that recover the visible world partially, but well enough to carry out a task (purposive, animate or behavioral vision), and we contrast it to the traditional paradigm of treating vision as a general recovery problem.

Co-authors

Bibtex Entry

@inproceedings{AloimonosDFHRS92i,
  title = {Behavioral Visual Motion Analysis},
  author = {Yiannis Aloimonos and Zoran Duric and Cornelia Fermuller and Liuqing Huang and Ehud Rivlin and Rajeev Sharma},
  year = {1992},
  booktitle = {IUW},
  pages = {521-541},
  abstract = {We propose here a new approach to addressing problems related to visual motion, namely the purposive approach. Instead of considering the various visual motion tasks as applications of the general structure from motion module, we consider them as independent problems and we directly seek solutions for them. As a result we can achieve unique and robust solutions without having to compute optic flow and without requiring a full reconstruction of the visual space, because it is not needed for the tasks. In the course of the exposition, we present novel solutions to various important visual tasks related to motion, such as the problems of motion detection by a moving observer, passive navigation, relative-depth computation, 3-D motion estimation, and visual interception, using as input only the spatial and temporal derivatives of the image intensity function. It turns out that the spatiotemporal derivatives of the image (i.e. the so-called normal flow) do not seem to be capable of solving the general "structure from motion" problem. They are, however, sufficient to provide robust algorithms for the solution of many interesting visual tasks that do not require the full solution, but only part of it. The ability to create robust nontrivial behaviors suggests the possibility that visual perception could be studied as intelligent behavior. We point out some of the benefits and drawbacks of this paradigm that studies vision as a set of behaviors that recover the visible world partially, but well enough to carry out a task (purposive, animate or behavioral vision), and we contrast it to the traditional paradigm of treating vision as a general recovery problem.}
}