Wide-Baseline Image Matching And 3D Camera Poseestimation Algorithm

In the case of wide baseline stereo matching, the goal is to find corresponding points that are projections, in the different views, of the same point in 3D space between images captured by different camera at different view point or same camera but different perspective posture. This is one of the basic problems in computer vision and pattern recognition which has not yet been completely resolved. Many problems in computer vision involve as a key subroutine solving a version of the correspondence problem. Whether the issue be properly addressed or not would have a certain influence on the performance of the algorithm.The topic of this paper comes from the international scientific communication and cooperation project (project number: 2008DFA12300). In order to solve the problem of simultaneous wide-baseline image matching and real time camera pose estimation, a fully affine invariant feature point descriptor is proposed. The method performs in three steps. First, a feature descriptor is built based on DCT decomposition of the template image and linearization of affine transform, and an affine camera pose model is quantized to made this descriptor affine invariant. This descriptor allows carrying out, in the second step, a rough pose estimate via descriptor matching. The last step is a fast verification made possible by optimization method which has been widely used in the domain of image alignment. Experiment results show that this descriptor outperform not only in speed but also in accuracy than the state of art affine invariant descriptors. Furthermore, it can achieve complete affine invariant matching and simultaneously three-dimensional camera pose estimation. Therefore, the descriptor can be used for real-time three-dimensional object matching, simultaneous localization and mapping (SLAM), object detection, etc.