Depth Information Restoration Of Three-dimensional Scene Based On Robot Binocular Stereo Vision

Stereo reconstruction of dense depth maps from natural video sequences is a fundamentally important and challenging problem in computer vision. The reconstructed depths usually serve as a valuable source of information, and facilitate applications in various fields. The problem of the image noise, textureless and occlusions regions exist in the stereo matching method, so it is difficult to get high-quality depths. In this paper, a novel method was proposed to automatically construct a view-dependent depth map for each frame. The corresponding depth values in multiple frames were consistent and distinctive depth values were assigned for pixels that fall in different depth layers.We introduced camera calibration and used two cameras added by a robot to acquire the inside parameters of cameras. The outside parameters for all frames can be estimated reliably by the structure from motion. Then we search methods to eliminate the image noise and make the depth estimate of both the textured and textureless regions can be effectively improved.After the parameters for all frames were estimated reliably, we first initialize the disparity map for each frame independently. Then we established a model of photo-consistency eliminate. Belief propagation based on Markov random fields was proposed to solve the maximum a posterior estimation problem. Multi-view stereo methods combine the algorithm of belief propagation. The photo-consistency constraint can solve the problem of adjacent area depth smoothness. In order to get an accurate depth information estimate in the textureless region, an improved color segmentation method was proposed.On that basis, we combined the photo-consistency and geometric coherence constraints associating some views in a global energy minimization framework. And we establish a unified framework which can reduce the estimation outliers possibility of all views depth information. They reduced the influence of image noise and occlusions acting on depth information restoration. We can produce sharp and temporal consistent object boundaries among different frames. This process is rather effective in estimating temporally consistent disparities while faithfully preserving fine structures.