Research On The Three-dimensional Reconstruction System Based On Multi-view Stereo For The Three-dimensional Swept Volume Display Three-dimensional Swept Volume Display

Three-dimensional (3D) displays are considered as the trend of the future display systems, and swept volume display is one of the latest. Comparing with the traditional two-dimensional (2D) display and binocular based stereoscopic display, it can provide more direct visual cues and can be observed by multi-user from multi-view simultaneously without using any special headgear or glasses.Multi-view stereo based3D reconstructed systems can obtain the voxel information of the real world objects, which will make the swept volume display to be widely used in the areas of military, medical, scientific visualization, manufacturing engineering, culture and entertainment, virtual commerce, education and training, etc. and to be a important rule in the future market of the3D display.This paper focuses on the multi-view stereo based3D reconstruction system for the swept volume display, mainly in three aspects:camera network, stereo correspondence and depth map fusion.For the camera network, we first analysis the common exist non-effective region phenomenon in the cameras, and design the mechanical structure of the network according to the360degree visibility principle of the swept volume display; then, we propose the virtual camera model to handle the object overflowing problem happened in the stereo rectification process. Results show that the object captured in the effective region of our network can be represent in360degree view, and the view of each camera is utilized maximally.For the stereo correspondence, we first analysis the challenges of improving the efficiency of the algorithm and the accuracy of the correspondence. The algorithm efficiency is limited by the lack of a suitable object segmentation approach for the common scenes, hence the uninteresting areas are time consuming for computing. The correspondence problems are the different light intensity of the same object from various view angles, the ambiguous boundary, sparse texture, repetitive texture. Then we propose our plan:we first segment the interested object from row-aligned images by optical flow, then use the Normalized Cross-Correlation(NCC) function based windows to overcome the various light intensity problems, direction window to fix boundary problems, statistical parameters to sparse texture issues, precise initial matching principle to repetitive texture troubles. In addition the image pyramids can be applied to the improving the computation speed in the cost of losing tiny image accuracy. The results show the object can be segmented and the common correspondence issues are successfully handled.For the depth map merging, we first give the formula for the single depth map calculating, then the coordinate system transformation between cameras are proposed, at last we provide our volumetric filtering method to overcome the reference factor during the merging process. The simulating results show that object the can be well represented in360degree view.A framework from camera networking designing to disparity map computing and depth map merging is proposed, which has theoretical and practical meanings to the areas of3D objection reconstruction and processing.