Design And Implementation Of Depth Estimation Algorithm For 3D Scene With Edge Constraints Triangulation

Depth estimation is one of the research hotspots in the field of computer vision.Similar to the perception ability of human vision for object depth,computer vision can calculate the distance from the point in the scene to the camera plane through image texture and multi view geometry.Depth information can effectively reflect the three-dimensional state of the scene.With the continuous development of mobile robots,autonomous driving and other related fields,the real-time research of depth estimation has been widely concerned by academia and industry.In this paper,we design and implement a binocular depth estimation algorithm which can process high-resolution image quickly by using the efficiency of triangular mesh.This algorithm improves the accuracy of the mesh by object edge extraction and mesh plane fitting and completes the depth estimation by constructing Bayesian generation model.At the same time,this paper also implements a depth estimation system based on mobile platform.The main contents of this paper are as follows:1)Binocular matching supports point extraction and parallel triangulation.In order to solve the problem of low efficiency of the current stereo matching algorithm for high-resolution images,this paper makes use of the efficiency of triangular mesh to improve the matching efficiency.In this paper,we screen the candidate points by fast Sobel operator combined with left and right consistency constraints and fuzzy matching constraints.Then the supporting points are triangulated by region overlapping partition and conquer triangulation method,so as to quickly construct the initial triangulation mesh of the image and lay a foundation for the subsequent steps.2)Triangle mesh optimization based on edge constraints.In order to solve the problem of large error on the edge of triangular mesh,the relationship between triangular mesh and object edge is studied.The cost function is constructed by the information of the normal direction,regional color complexity and maximum gradient intensity of each triangle in triangular mesh,and the judgment of whether each triangle crosses the object edge is made by combining the MRF modle,and the triangular mesh that crosses the object edge is screened out.Then,the grid is adjusted quickly by the triangular grid constraint algorithm in the minimum influence domain,so that the grid can describe the environment contour more accurately.3)3D mesh optimization and local interpolation algorithm based on multi plane fitting.In order to solve the problem of mismatching of mesh vertices,a multi-plane fitting algorithm based on triangular mesh is proposed to optimize the mismatching points in some weak texture planes.The algorithm makes full use of the structural characteristics of the triangular mesh and the prior edge of the extracted object,so that it contains 3D points along the mesh edge,which has faster speed than the traditional plane fitting method.Then a dense depth map is obtained by interpolating a robust triangular mesh with a bayesian model.4)The depth estimation system based on mobile platform.In this paper,the algorithm is parallelized on the mobile platform and a robot platform system is built autonomously.The triangulation,pixel operation and triangulation operation are optimized in parallel,so that the system can run in real time on the mobile platform.In summary,this paper builds a priori of depth by generating a triangle mesh on a set of support points which can be stably matched.Then we extract the priori information of the object edge by modeling the triangular surface information.And we dynamically adjust the two-dimensional mesh structure and optimize the depth of three-dimensional mesh,so as to deal with the problems caused by the occluded boundary and the weak texture plane.Finally,we get dense disparity map by interpolation.This algorithm can effectively reduce the parallax search space to effectively process high-resolution images.In addition,since the algorithm can quickly determine the parallax range,it is easy to realize parallelization,which is important for the algorithm to realize real-time performance on embedded devices.