Research On Depth Information Estimation For Computer Vision

As a multi-disciplinary field,the theoretical study and the practical application of computer vision have developed rapidly.The research on the depth as input information of the visual information system has become a popular topic in computer vision.How to estimate the depth information from a two-dimensional image effectively has become a key issues of the research.The depth information can be estimated using a single or multi-view image.The extracted depth map can not only generate three-dimensional images but also achieve image-based rendering and the reconstruction of three-dimensional model.It has become the basis of computer simulation of human vision.With the wide application of the relevant technology,the depth information estimation accuracy,the computational efficiency,and the ability to handle complex scenarios have demanded a higher standard.This paper explores the key issues of the depth information estimation using a single viewpoint image and the binocular stereo matching.(1)To overcome the high complexity and the large computation of the recovery scenarios depth information algorithm using image high-level cues,this paper presents a depth information extraction algorithm for the single view image using profile sharpness of image low-level cues.The algorithm takes the profile sharpness information of edges as the estimating characteristics of the blurred information.It establishes an improved model of contour tracking with the edge contour sharpness information,and consequently the contour can be extracted.A prior hypothesis of depth gradient is used to assign depth to extract the scenario depth information,and the depth map is optimized by the cross bilateral filter.Experimental results on a variety of images show that a liable extraction of the depth for the single view image can be acquired availably with this simple but effective algorithm.In order to enhance the utility of the algorithm,this paper presents a FPGA-based IP core design.(2)To reduce the mismatching of the classic Census Transform at the depth discontinuity as well as under the noise disturbance,this paper proposes a modified Census Transform based on information of the neighborhood for stereo matching.This new Census Transform utilizes 2-bits to represent the differences between a pixel and its neighborhood.The result image of the transform provides more details at the depth discontinuity,and effectively distinguishes differences between the center pixel and its neighborhood pixel located at the weakness texture region and the depth discontinuities region.It also minimizes the impact of noise on the quality of matching.The stereo matching algorithm based on the modified Census transform calculates the initial matching cost by a sparse computation of the Hamming distance.The weighted cost aggregation of the algorithm improves the matching accuracy.After the sub-pixel interpolation,left-right consistency check and the interpolation of the occluded regions,a dense disparity map can be obtained.The evaluation of Middlebury Stereo images shows that the proposed algorithm enhances the accuracy of matching with concise structure and lowers the complexity with strong robustness.To improve the practicability of the algorithm,this paper presents a FPGA-based parallel hardware acceleration program of the algorithm.With the parallel technology and pipelined design methods to enhance the handling capacity of the system.(3)The cost aggregation is computationally expensive with low speed,which seriously affects the real-time performance of stereo matching algorithm.This paper proposes a hierarchical cost aggregation algorithm that divides the cost volume into multiple levels by employing pyramidal decomposition and aggregates the costs using the adaptive support weighting cost aggregation at each pyramid level.In order to improve the recovery accuracy at the depth discontinuity region,this hierarchical cost aggregation algorithm adopts an edge-preserving piecewise linear approximation up-sampling methods to achieve cost aggregation recovery from a coarse level to a fine level.Experiments show that the algorithm reduces the computational complexity and provides good characters.A hardware implementation architecture for the adaptive support-weight approach of cost aggregation is designed to improve the traditional adaptive weighted aggregation algorithm using weights approximate to reduce the computational complexity.During the hardware implementation,it replaces the multiplication operation with the shift operation,and uses a pipelined cache mode and dual-channel cost aggregation method to speed up the cost aggregation.(4)To handle the problem of high erroneous results of the stereo matching occured at the depth discontinuity region,the slanted surface and the non-front-parallel surface,this paper presents a stereo matching algorithm using an improved Patchmatch and a slice sampling particle belief propagation.This algorithm establishes a model of the depth estimation for the non-front-parallel surface which using an edge-preserving similarity function of the Patchmatch.It replaces the nearest neighbor search with the particle belief propagation and approximates the target distribution with a finite set of particles.The sampled particles from the belief distribution is typically done by using slice sampling Markov chain Monte Carlo methods to solve the particle updating problem.The method can improve the effectiveness of the resampling particles and enhance the matching accuracy.The experiments show that the proposed algorithm reduces the mismatching at the depth discontinuity region effectively and improve the accuracy of matching for the slanted surface and the non-front-parallel surface.