Research On Depth Estimation Algorithms Based On Monocular Video

With the development of digital signal processing, people's life has changed a lot. And, our hearing and seeing is going through a revolution. As the primary tool for view extending, television has gone through several revolutions. The future revolution will be three dimensional television - Stereoscope Television.Stereoscope television is a three dimensional displayer without any assistant tool, and people can get the three dimensional image from any point of view. At present, this displayer has appeared in the market, such as Philips' 3D displayer, and Sungsum's 3D displayer.As the 3D video format, composed of two dimensional (2D) video and depth map is considered as the most popular one. This thesis discussed the depth map generation algorithms, i.e., 2D to 3D video conversion. During the processing, manual image processing is used for key frame's depth map generation. This research belongs to semi-automatic 2D to 3D video conversion.Two general semi-automatic depth map generation algorithms are discussed at the beginning, e.g., depth map generation based on video segmentation and motion estimation, and depth map generation based on Bilateral filter. A depth map generation algorithm based on foreground area segmentation using bi-direct Bilateral filtering is proposed at the end of this thesis.For depth map generation based on video segmentation and motion estimation, the accuracy of motion vector is vital for the result. Additionally, block artifacts are always involved since the seminal color for new occluded area and moving object. Here, an improved algorithm based on the initial depth for key frame is proposed. This can be considered as an adaptive block matching algorithm for motion estimation. It improved the accuracy of motion vector, as well as computing complexity.For depth generation based on Bilateral filtering, to supply the depth variation during moving, a bi-direct Bilateral depth generation algorithm based on belief map is proposed. And, the simulation results tell that it is effective for new color areas.In one word, depth map generation from monocular 2D video is a new research topic. In this thesis, an improved algorithm for semi-automatic depth map generation is proposed, and the visual results show that the proposed algorithm is effectively in depth map generation.