Multi-scale descriptor for sequence image based vision applications

Abstract

Finding correspondence points is a fundamental problem in robot vision applications such as object recognition, 3D reconstruction, and camera motion estimation. There are many problems to be solved. One of the typical problems is finding correspondence points under scale, rotation, illumination, and viewpoint changes between two images. Many proposed local descriptors such as SIFT and SURF descriptor have showed good performance under rotation, scale, viewpoint, illumination changes conditions. In sequence image based applications such as camera motion estimation, object tracking, and visual SLAM and etc, a descriptor that is invariant to large scale or rotation changes is not essential. Since, these appli-cations find correspondence points with small scale or rotation changes. In sequence image based vision applications, the most important properties of the descriptor are its distinctiveness for only small rotation and scale changes and low computation complexity. This dissertation focus on a novel descriptor that is suitable for sequence image based vision applications. We present a novel descriptor that requires low computational complexity and offers high precision. The proposed descriptor achieves a high recognition rate through the use of multi-scale information instead of single-scale, as combing the each scale descriptor leads to improve the representation of characteristics of interest points. The proposed descriptor provides high robustness but slow to compute and match because of its multiple descriptors. In order to reduce computational complxity of multiple descriptors, the proposed descriptor is calculated by employing 4 global integral images and an intensity gradient. We also propose 2-stage cascade matching method to find correspondence points more efficiently. We evaluate the proposed descriptor in comparison with different types of descriptors using benchmark datasets in the chapter 3. The results show that the proposed descriptor provides ...