Computational Vision in Bad Weather

Columbia University, New York NY

Investigators

Abstract

Computational vision has made significant strides in the development of sensors and algorithms that recover scene properties from images. However, virtually all work in vision is based on the assumption that the observer is immersed in a transparent medium (air)and that the objects of interest are opaque. It is assumed that light rays reflected by the objects travel to the observer without attenuation or alteration. Therefore, existing vision sensors and algorithms have been created only to deal with clear weather. In practice, however, a vision system must reckon with the entire spectrum of atmospheric conditions commonly known as bad weather. It must continue to perform in the presence of a variety of conditions,including, haze, fog, rain, hail and snow. This proposal outlines a comprehensive research program geared towards the development of models and methods that can aid vision in bad weather. The first step is to understand the visual manifestations of different types of weather conditions. For this, we will draw on what is already known about the optics of the atmosphere. Since the atmosphere modulates the information carried from a scene point to the observer, it can be viewed as a mechanism of visual information coding. We propose the development of a general computational framework that exploits the brightness and color changes that are induced by bad weather. Based on this framework, we will develop models and methods for recovering pertinent scene properties such as true (clear weather) color and three-dimensional structure, from images taken under different (unknown) weather conditions. Such models and methods have obvious applications in scene understanding, autonomous navigation and video surveillance. In addition, we wish to create an extensive image/video database that captures the wide range of visual effects caused by weather. We believe that such a database will make it easier for researchers to study this important area of computational vision.