A Self-adaptive ASIFT-SH method

Peter Podbreznik

doi:10.1016/j.aei.2012.11.001

2013-01-01

PRE

Peter Podbreznik *

Božidar Potočnik

DOI：10.1016/j.aei.2012.11.001

原文链接

原文求助

摘要

En 中文

When monitoring events on a building site using a system of multiple cameras, it is necessary to establish correspondence between the acquired imaging material. The basic problem when attempting this task is the establishment of any correspondence between points located on uniform areas of the images (e.g. regions with uniform colour or texture). The basic version of our ASIFT-SH method can mainly solve such a problem. This method consists of four steps: (i) determining the initial corresponding points within the images of both views by using the ASIFT method, (ii) grouping of initial corresponding points from the first step into subsets, based on segmented regions, (iii) calculation of local homographies for a particular subset of corresponding points, and (iv) determining any correspondence between arbitrary points from a particular camera's viewpoint, by using a suitable local homography. The critical step of this method concerns segmentation. Therefore, we have introduced into our algorithm a step for adaptive adjustment, the segmented regions being remodelled so that they better meet the required coplanarity criterion. This introduced step is based on a 3D reconstruction of the initial corresponding points and a search for the minimal number of planes within the 3D space to which these points belong. Those points that belong to a particular plane, represent a newly-created subset of the initial corresponding points. The results point out that the introduction of this adaptive step into ASIFT-SH significantly improves the accuracy of corresponding points' calculation. The mean error is 1.63 times lower and the standard deviation is 2.56 times lower than by the basic version of the ASIFT-SH method. (C) 2012 Elsevier Ltd. All rights reserved.

Keyword:

Corresponding point

Arbitrary-selected point

Widely-separated views

3D reconstruction

Segmentation