Augmented Reality(AR) is a challenging topic in computer graphics and computer vision, in which AR techniques for small nature scenes are widely used in military, manufacturing, medical, entertainment,advertisement and cultural construction. This paper focuses on small nature scenes AR techniques, mainly discuses about automatic initiation methods, robust augmentation technique under illumination change and partial occlusion. Our contributions are: 1) We present a complete geometry reconstruction method from two views. In computer vision, given two images and corresponding cameras' inner parameters and poses, depth of points in 3D scene can be obtained by solving equations. Being different from solving equations, this paper presents a complete geometry method for 3D points' depth estimation. Its key idea is eliminating the parallax caused by rotation between two cameras, and then breaking down 3D translation to 2D move. The distance of 3D points to camera is obtained by similar principle of triangles in 2D planes. Experiments show that geometry method gives as satisfactory reconstructions results as those of least squares on precision. In some applications such as embedded development for robot stereo vision, our method has advantage for its computing and storage costs are economical. 2) A real-time framework for objects cursory recognition in cluster scene based on visual attention is introduced. First, multi-scale image features are combined into a single saliency map. Then, k-means method is used to estimate the position of objects from cluster scene by saliency map. Finally, we construct global color feature vector for saliency regions and recognize the objects by their correlation coefficients from patterns. Results show that this framework is efficient for objects cursory recognition in random cluster scene. 3) Outlier rejection is a challenging topic in computer vision, especially where illumination change or partial occlusion happen in real-time image matching. This paper presents a mechanism for reliable match selection from ambiguous match which are obtained by invariant feature operator. Its basic idea is that if one pair is a good match, then many matches will be expected to see in their neighbors. We use this mechanism for high reliable point selection and compare our results with those of classical method. In the last part of this paper, we combine our method with RANSAC for outlier rejection in point match. Experiences show that our met...
修改评论