Comprehensive comparison of gradient-based cross-spectral stereo matching generated disparity maps

In Gradient-Based Cross-Spectral Stereo Matching (GB-CSSM) output disparity maps tend to produce coarse results that are, for the most part, reliable. However, general methods of improving the performance of disparity maps generated from the Cross-Spectral comparison of visual and full infrared input images are non-existent. In particular, previous works fail to address the role and interaction of the parameters present in a GB-CSSM algorithm as a way to improve the performance of a disparity map. In this paper, we introduce the first comprehensive comparison of GB-CSSM generated disparity maps. More specifically, we consider all possible input parameter combinations to a GB-CCSM algorithm and evaluate how these parameters affect runtime as well as accuracy and validity of the disparity maps. Our objective is to provide designers with a systematic means of classifying and easily identifying optimal disparity maps (in terms of a combination of runtime, accuracy, and validity). Our experimental results show how a Pareto frontier of optimal disparity maps can be generated as a result of our analysis. The ultimate goal is to allow for the development of new and improved GB-CSSM algorithms that can be applied to a broader range of applications.

I. Introduction

Visual stereo matching is of limited value when weather conditions offer zero visibility. Under these circumstances, visual image capture is not possible. An aircraft performing search and rescue over an area of interest is therefore constrained to use full IR cameras in an attempt to match as accurately possible the area of interest against a visual image of the same area taken at an earlier date when weather conditions were favorable.