A Study Of Interactive Segmentation And Object Location In Image Registration Of Video Scene And Virtual Image

Image segmentation and object tracking are important research task in the field of computer vision, they are wildly used in areas such as pattern recognition, image editing, intelligent surveillance. We use image segmentation and object tracking to realize the image registration of true video scene and the virtual image. For more vivid effect, we propose a whole completed solution with three steps such as using interactive segmentation to gain the boundary of object, estimating 8-parameter of the perspective transformation, image registration by high resolution resampling.In the step of interactive segmentation, this paper expatiates and improves the Graph Cuts method based on Max-flow algorithm. By watershed and region merging, also some method of reducing the valid segmentation range of image, we construct the flow network less complicated for accelerating the algorithm.In the step of parameter estimation, this paper labor the method of using the Levenberg-Marquardt algorithm to estimate transformation parameter by minimizing the cost function based on brightness invariant, meanwhile the "coarse-to-fine" strategy is adopted for computational reason and avoiding incorrect matching. After that, we introduce an improved LM algorithm to reduce the computational magnitude. Besides that, we propose a novel object tracking algorithm based on Kalman filter, Mean Shift and perspective transformation. Experimental results show that the algorithm is reliable and efficient even in the circumstance of fast object moving.In the end this paper adopt a novel image registration method based on high resolution resampling, meanwhile consider the effect of light, we generate the shadow for the virtual input image and make it in the area of the object's boundary which obtained in first step, then the final effect of image registration may as real as possible.