Research On Depth Enhancement Algorithm

Depth map, as a method to record 3D information of the scene, has been widely used in computer vision, robotics and human computer interaction. Different from texture image that can be captured with high quality, acquiring accurate depth map has been an open issue. In depth map, the interference problem and the large-area depth missing problem are two main serious problems. This paper mainly focuses on such two problems.For the interference problem in the multiple depth sensors system, in this paper, hierarchical patterns are proposed by analyzing measurement principles of the structured light technology. In the hierarchical pattern, both de Bruijn sequence and pseudorandom array are applied. The de Bruijn sequence is used to ensure that each encoding unit is unique in the horizontal direction, while the pseudorandom array is adopted to guarantee that each pixel in the encoding unit is unique. By designing different de Bruijn sequences and pseudorandom arrays, multiple orthogonal hierarchical patterns are generated. For depth calculation in the interference problem, different patterns are separated by analyzing their orthogonal properties. Then, coarse depth map is obtained by decoding de Bruijn sequences and accurate depth map is generated by decoding pseudorandom arrays.For the large-area depth missing problem on planar surfaces, in this paper missing depth values are recovered via ProSAC. Considering specialties of the problem, namely that almost entire depth values are missing and only contour pixels are available, the ProSAC algorithm is applied to select high-quality contour pixels for surface function approximation. Missing depth values can be recovered by referring to the approximated surface function. In the quality assessment of each contour pixel, both the texture similarity calculation and the histogram analysis methods are applied. Since the histogram analysis method is independent from texture information, the proposed method has better applicability.Experimental results further demonstrate that the proposed approach can effectively improve the accuracy of the depth map, even when there exists the interference problem or the large-area depth missing problem.