Research On Scene Topology Combined With Scene Depth Estimation

The research of scene depth structure of monocular images has always been an important part of current computer vision tasks.This task focuses on how to obtain the corresponding scene depth structure from monocular images.In the computer vision theory established by Marr,this is an important task for human vision.The scene depth structure generally refers to the topological structure relationship in the three-dimensional space formed by different objects in the scene under the same observation device and this relationship is reflected in the monocular image as a two-dimensional planar relationship.How to obtain the front-to-back ordering relationship of the corresponding target from the monocular image is the final problem to be solved in the scene depth topology.There are still several challenges in the current depth ordering task of depth scenes in monocular images:(1)It is difficult to obtain depth clues in complex scenes with cluttered objects.The disordered objects and the suspension of targets directly affect the detection of vanishing points,leading to the lack of depth clues;(2)Occlusion exists as an important clue to restore the depth relationship of the target area,but in some complex scenes,there are often cases where there is no direct occlusion relationship between some objects,and even some occlusion relationships conflict with those in real scenes.In response to these issues,the following work has been carried out in this article:(1)This topic summarizes the current research status of deep feature acquisition and analysis,and fully elaborate on the key steps of scene depth estimation,public data sets and evaluation standards.The topic explains the characterization method used in scene depth information,and explained the basic theory of existing methods from three aspects of non-deep network,deep network and graph model,which laid a theoretical foundation for scene depth estimation and topology structure research.(2)Aiming at the problem of missing clues caused by the detection of vanishing points in the target scene,this thesis uses pixel-level depth estimation to replace the corresponding vanishing point clues,and builds an U-Net network based on atrous spatial pyramid pooling.The function of expanding the receptive field is realized without losing the resolution in the encoding stage,and the addition of multi-scale cascade operation avoids the occurrence of information redundancy.In order to take into account the depth estimation of objects at different scales,the model usessampling across layers to collect scene feature information of the same scale,which effectively improves the model's convergence accuracy.Our experiments on the NYU-Depth-V2 dataset,which is a well-known scene depth sorting,prove the advantages of this method.(3)Aiming at the effect of occlusion judgment failure or conflict with occlusion relationship on the depth ranking results,this thesis proposes to use the regional occlusion results as local depth cues and combine the depth estimation information as global depth cues to obtain the final depth ranking results.In order to effectively integrate local depth cues and global depth cues,this chapter uses Hidden Markov Model to integrate the two cues,constructs a graph model by using two types of deep cues,obtains relevant model parameters and inferences based on EM algorithm parameter learning,and gets the final depth sort result.Experiments on the recognized scene depth ranking NYU-Depth-V2 and Make3 D datasets prove that the method proposed in this thesis can effectively improve the research of scene depth topological ranking,and is better than the current mainstream scene depth ranking methods in terms of corresponding errors and accuracy.