Research Of Wide Baseline Stereo Matching Based On Invariant Feature

Computer vision mainly studies how to realize human visual function with computers . The main idea is to perceive , identify and understand the three-dimensional scene from two-dimensional projective images.Stereo matching is one of the key issues in computer vision.The information of depth or distance can be obtained from disparity of the points.Because of the larger angle deviation and the larger impact of occlusion and light in the condition of wide baseline, wide baseline stereo matching is an ill-posed problem.How to extract affine invariant features from images is a common problem in many fields such as object or scene recognition, scene matching,image retrieval,etc.Compared with other features, affine invariant features are invariant to the viewpoint changing and camera parameter variance,and have great advantage over robust, distinction and applicability,hence the theory and method of affine invariant feature extraction have became an challenging field.Although a variety of works have been done in the field, wide baseline stereo matching is still a challenging problem. This thesis is focused on some practical issues on wide baseline stereo matching.The major works of this paper can be generalized with three aspects:1. Images from different perspectives have the bigger perspective, rotation, scale and lighting changes in the condition of wide baseline stereo matching. They cannot directly use the correlation algorithm based on gray matching or reconstruction. In the paper, it used scale space theory to extract invariant features.2. In the scale space, the dimensions of descriptors are large changes and the behind operation is not efficient. It was using the ideas of Principal component analysis to reduce the dimension of descriptor and improve the matching efficiency.3. It has the higher mismatch of wide baseline stereo matching in the condition of occlusion. In this paper, the image which is on the right and left sides was respectively as reference image. It detected occlusion areas and remove the mismatch which was caused by block.The experimental results show that the improved scale invariant feature transform algorithm can perform well. In the condition of wide baseline, especially a larger perspective changes, illumination changes and larger scale changes. The algorithm can still find the matching area between two images, so it has a certain reliability and robustness.