Tri-Stereo Image Analysis

Abstract

In computer vision the term shape from x is used which investigates the extraction of shapes from an image using x cues, where x is shading, photometric stereo, focus, defocus, texture, motion, and stereo. All these methods investigate the images to generate a three dimensional view using single image with known light source geometry or stereo pair with known camera geometry.

The stereo image is individually a 2D dimensional representation of the real world. The image provides a method to generate a depth map using image correspondence, but the problem is that stereo matching is an ill posed problem and hence there does not exists a unique solution. Hence stereo images can be matched using optimization processes. The transformation between image space geometry and the real world object

space which relates with the 3rd dimension of the real world objects is achieved through relating these two spaces using photogrammetric principles and Ground Control Points. For analysing Lunar stereo images obtained from Terrain Mapping Camera (TMC) used on Chandrayan-1 mission, it is not possible to define such GCPs, hence a novel attempt was made to obtain stereo in Trinocular mode, whereby it would be possible to generate three relatively oriented stereo models of object space and then a optimization of these 3 Independent models could be used to dense optimal 3D model of the Lunar terrain. In my PhD, the research is planned to develop this optimization and in the process some novel techniques to improve the working of interest point extraction and trinocular matching are also being attempted.

Interest point operators that identi.es an image points based on its gradient properties with its neighbouring pixels is by far the most popular technique used here. From literature it can be observed that most of the commercial implementations use Frostner operator in some form to identify conjugate points in a stereo pair. In recent times many other algorithms are proposed such as SIFT and SURF, that are generic and not speci.c for

stereo correspondence, but indicate promise to improve the reliability and speed of point extraction over the Frostner operator. In case of dense matching the results form feature matching can be used as an input to get the a prior but later on area based methods are necessary in particular to remote sensing applications. Some of the preliminary results are presented here.

Stereo matching is an active research area and there are many methods proposed for image correspondence, but still for lunar missions the classical correlation based methods are used. The research will investigate the use of other algorithms for stereo matching. The TMC operates in triplet mode thus giving again a challenging task to relate model with diff.erent camera con.figuration.