|M.Sc Student||Eran Pinhasov|
|Subject||Optimal Usage of Color for Stereo Vision|
|Department||Department of Electrical Engineering||Supervisors||Full Professor Shimkin Nahum|
|Professor Emeritus Zeevi Yehoshua|
Stereo vision finds the depth of image pixels (dense stereo) or key locations (edge-based stereo) using two or more scene images taken from different view angles. Digital color imaging systems are very common today, still, most of the stereo algorithms use only gray level intensity images, like the Y channel from the YUV space, assuming that the Luminance channel holds most of the information, while the other chromatic channels are redundant. In this work we propose to improve the disparity estimation by working in a transformed color space, which is created by taking linear combinations of the input multi chromatic image channels (Tri-Color in most cases). The optimal color space is found by minimizing the disparity estimation variance, which is computed from the left and right images data. The calculated variance expression has the form of a quotient whose numerator depends on the noise covariance matrix and the denominator depends on the horizontal derivatives cross correlation matrix of the noise free left image. The transformed images can be used as inputs for existing stereo algorithms. We examine both local and global versions of this process, as well as other extensions that include a local adaptive approach, which selects the number of channels to use during the correspondence process (locally) depending on the calculated disparity error variance. We present several methods for estimating the noise covariance matrix and the derivatives cross correlation matrix, which are needed for finding the optimal color space. The improved performance of these methods is demonstrated using synthetic and real stereo pairs.