Research On Scene Depth Estimation Of Combined Occlusion Detection

The depth estimation of monocular scene images is used in many visual tasks such as3 D scene reconstruction,visual navigation,image segmentation,and human pose estimation.Its purpose is to obtain the scene depth information of an image from a two-dimensional image.Existing deep learning methods are still interfered by scene ambiguity information and noise information during the determination of local depth and scene depth,making it difficult to extract complex depth patterns.This is because learning the 3D information in the 2D scene itself is morbid,for a infinite number of 3D scenes can be mapped to the same 2D scene picture.As a result,the depth estimation task has always been a difficult task for computer vision tasks.The paper subdivides scene depth estimation into two sub-problems,and researches on partial occlusion relation and scene depth estimation respectively.The main workload of this thesis is as follows:1 The subject elaborates the research status of scene depth estimation from physical inspiration to deep inspiration,and explains the related method and process of physical inspiration and depth inspiration respectively,which provides a theoretical basis for extracting partial edge occlusion and scene depth estimation using depth model.2 In view of the fact that occlusion edge detection is easily interfered by noise information and inspired by the common features of the scene edge occlusion relationship and edge detection,the subject proposes a multi-branch model with a shared backbone network for extracting edges and edge orientation detection at the same time.In this model,edge detection and edge orientation detection use a shared backbone network,and merged by non-maximum suppression,which can suppress the interference of non-edge information and extract the occlusion relationship representation on both sides of the boundary.On the PIOD data set,the accuracy of the proposed method in edge detection and edge orientation detection is verified.3 Aiming at the problem of model degradation caused by deep estimation feature learning that is susceptible to vanishing gradient,the subject proposes a deep network design that combines residuals and dense connections.First,the down-sampling operation in the convolutional network is used to collect the multi-scale features in the image;then,a Residual Dense ASPP module is proposed to extract the depth pattern of the complex scene;finally,the feature resolution is restored through the up-samplingprocess to complete the high-precision scene depth estimation.The experimental results of the NYU-Depth-v2 data set on the accepted scene depth estimation data show that higher accuracy and smaller errors are obtained than other current methods.The method proposed in this thesis can improve the problem of inaccurate object boundary prediction in complex scenes.