Research On Quasi-dense Matching Between Fisheye Images For Structured Scenes

In computer vision, the matching of image point correspondences is one basic problem, which is also a core problem. For the most computer vision problems (such as motion estimation, object recognition, target detection, and 3D reconstruction), establishing the point correspondences is the most core step. So far there are many matching algorithms available which could be roughly divided into three categories, named sparse, dense and quasi-dense approach respectively. Conventional quasi-dense matching algorithms are designed for general scene structures, which have not considered the structural scene contains large planar structures. And most researches about the matching of image point correspondences are focusing on normal perspective images. This paper proposes a quasi-dense matching method based on homography, and proves its better quality than some other methods. The main points in this paper are as follows:1. A novel quasi-dense matching method is proposed for structured scenes, which is based on plane induced homography. First, sparse planar seed regions are detected and matched by using affine invariant detector and matcher. Then region growing algorithm is employed to propagate seed region to its neighboring pixel by using the constraints of both planar homography and similarity measure. The best-first strategy is used which make our algorithm robust to initial sparse match outliers. Further more, more accurate reconstruction can be achieved by post-processing of unreliable matches. Experiment results show that our algorithm is more stable and accurate than conventional algorithms.2. A novel clustering method which is based on plane induced homography is proposed. The homography by this clustering method is more accurate than it calculate by best-first method. Because the seed planes which come from the same plane have the same homography. First, we can let them clustering in one group, then use the information of the seed areas in one group, calculated the initial homography. At the same time, threshold of the distance between the regional effect of clustering is given, and the clustering effect also gives the measure of the evaluation criteria. The initial homography which is accurated by best-first and clustering method is respectively used to region growing algorithm. Experiments results show that homography by the clustering is few iterations, faster growing, and more stable than the best-first.