Multi-View Computation Based On Geometric Invariants

With the rapid development of computer intelligence, computers are able to see things just like human being. They input the images captured by cameras and then perform diverse processing of the images; this is how computer vision, one of the most popular research domains recently, comes to existence. In compute vision research area, image matching, object recognition, etc. gain much concentration. For instance, in active vision system, intelligent analysis of objects-of-interest (OOI) in 3D scene generally needs multi-agent cooperation, which depends heavily on the accuracy of multi-view geometry computation and the robust correspondence of OOI. Thus, multi-view image matching algorithm becomes the vital part of the whole system. In most cases, active vision system captures multiple views under wide baseline stereo model, which results in distinct affine distortions and complex description of correspondence problem. Furthermore, the difficulty of depicting correspondence problem in active vision is aggravated when the cameras perform pose adjustment for active OOI tracking with the best viewpoint. It is known the geometry constraint between views is important to make correspondence problem well-posed. However, the multi-view geometry is time-variant in active vision because of the cameras'pose adjustment.In order to meet the needs of active vision system, the paper refers to geometric invariants, which describe the essential features based on pixels of images. Geometric invariants stay unchanged after diverse image transformations such as affine transformation, and they also have themselves carried by detectable features. Both of these two conditions ensure that image matching algorithm can perform well using geometric invariants. The paper gives a thorough and deep research on both wide-baseline image matching and correspondence of OOI.Firstly, the paper presents a comprehensive review on necessity of geometric invariants application and existing methods and then gives a description on suitable geometric invariants for wide baseline image matching.Secondly, the paper gives a detailed discussion on the whole process of wide-baseline image matching algorithm based on geometric invariants, including feature detection, feature description, feature matching and multi-view geometry generation. The paper also proposes an innovative method of image matching by combining one of the best feature detectors MSER and one of the best feature descriptors SIFT. The experiments demonstrate that our method can obtain more robust matching results between wide baseline images.Finally, the most striking creation of this paper is to propose a new solution to OOI correspondence problem between wide baseline images, named multi-epipolar-line constraint. It defeats existing methods by generating the more accurate one-to-one OOI correspondence and thus can be more suitable for real-time system. The experimental results show that our new method is very portable, robust and efficient.The paper also gives a prospect of the research related on multi-view computation based on geometric invariants.