Understanding And Application Of RGB-D Visual Content

Compared with 2D visual content,3D visual content contains more particularity and challenge. RGB-D image is one of 3D visual content, and it contains color information (RGB image) and spatial information (depth map). Depth data provide new clue for the research of 2D image. Compared to the point cloud data, depth data has low data volume, however, it also has low accuracy. So it is suitable for detecting and tracking in small scene. In recent years some RGB-D image acquisition devices with low cost and good robustness come out such as Microsoft Kinect, which makes researcher get RGB-D image easily. Since the research and application of RGB-D become more and more important, how to obtain the depth map quickly and accurately using existing images has become a worth problem to study.This paper explores the way to obtain depth information of image collection based on research of it. In a set of images of different perspectives on the same scene, these images have semantic similarity, we can think of their depth distribution is also similar.Firstly, SIFT Flow algorithm is used to obtain depth information of a single image. We input a set of images, and use the GIST features and KNN algorithm to retrieve candidate images from the RGB-D database which are similar with the input images. The SIFT flow algorithm is used to match the input images and candidates. SIFT flow use SIFT descriptor to match pixels in different images refer to optical flow. A three layers pyramid of classical SIFT Flow Algorithm is established by down-sampling to deal with the large computation problem, we find a rough position in the coarse layer and then refine the position in the accuracy layer. After matching, we can transfer the depth from candidates to the input, and estimate depth map of the input images.On the basis of depth information of a single image, image collection is used to refine depth data.In order to select the most accurate data from the candidates to make the depth image smooth, we use the space regularization method adopted by iterative calculation to refine preliminary estimation. Thesis uses homogenized Harris corner as a point of interest in feature collection. Both Haar wavelet transform and KNN algorithm are used to find the overlapping region of the image. The overlapping region is refined by Bilinear Interpolation algorithm. Experiment result shows that this method is effective partly.