Research On Feature Correspondences For Omnidirectional Images

Omnidirectional Vision, because of the large field of view, has become more poplar in variety of applications, such as robot navigation, video surveillance, virtual walkthrough based on real scene images, measurement and 3D reconstruction based on images and so on. At the same time, the hybrid vision system, which consists of omnidirectional and perspective cameras, is also used more widely than ever, since it takes advantage of a large view scope from omnidirectional cameras and higher resolution from perspective cameras. However, the complexity of the project model and the high distortion of the image challenge for many existing computer vision techniques, including the technique of feature correspondences. This paper mainly addressed the feature matching problems in omnidirectional vision and hybrid vision systems. The main points are as follows:1. Using fisheye images as an example, we proposed a feature matching method for omnidirectional images by introducing a conception of virtual camera planes. With projecting the image points to the virtual camera planes, the nonlinear distortion could be removed and many traditional matching methods for perspective images could be directly used on the corresponding problems between omnidirectional images.2. Based on the feature matching, we developed an automatic panorama mosaicing system with high distorted fisheye images. To increase the number of correspondence and improve the precision of the estimation of motion parameters, guided matching strategy is applied. By this strategy we got a motion model more accurately. Experiments showed that our system can lead to a stable result and has its generality.3. We present a solution for locating perspective images in omnidirectional images. To remove the huge number of outliers, which results from the different modes of omnidirectional and perspective images, we introduced a spectral technique. After the outliers are removed, we could estimate an approximate transformation between the perspective image and omnidirectional image, and locate the boundary of perspective image in the omnidirectional image. Then using the conception of virtual camera planes, we give a feature matching strategy in hybrid systems.