Multi-perspective Video Mosaicing With Sparse And Rotatable Viewpoints

Multi-perspective stitching can obtain a complete image of a long scene which is difficult to be captured by a single camera. It has been widely applied to remote sensing image processing, medical image processing,3D reconstruction, etc., and arouses the interests of researchers from computer graphics, computer vision and virtual reality in recent years. There are two limitations among existing algorithms:1) Numerous viewpoints are needed to extract enough strips for stitching; and2) each viewpoint is required to simulate facing towards to dominant scene plane or each viewpoint allows rotating a small angle for minimal strip distortion. This paper presents the following algorithms for overcoming these two problems:(1) A multi-perspective image mosaic algorithm with sparse and rotatable viewpoints.There will be different degrees of deformation in different parts of image when the optical axis of the camera is not perpendicular to the scene surface. Consequently, the neighboring images cannot be stitched together. We proposed a two-step stitching algorithm. The first step, virtual transformation, eliminates the different perspective distortions in different parts of the neighboring images, the second step, synchronous rectification and registration, updates the changed overlapping areas during rotation and thus overcomes the registration problems caused by the camera rotation.(2) An adaptive image resizing based image stitching algorithm.The image may show an unwanted aspect ratio after virtual transformation and cause the objects in the image with a tensile or compressive distortion. Therefore, an adaptive image resizing algorithm is proposed. It eliminates the tension or compression distortions by correcting the different regions of the image with a selected standard reference strips. To remove the ghosting effect caused by the depth variation after registration, we use a seam-line based strategy to obtain a more prefect mosaic.(3) A CUDA-based optimization method for fast stitching. The algorithm in this thesis processes each pixel in the image and, therefore, it is slow when only running on CPU. Therefore, we adopts a CUDA-based optimization method to reach fast computation. In the method, the virtual transformation process is transplanted from CPU to GPU. Experimental results show that this method improves the stitching speed and can obtain a real-time video stitching.