Robust Depth Estimation Techniques In Computer Stereo Vision

The visual ability of machines to see and perceive their environments in three-dimensions(3D)and distance is currently at the heart of computer vision.The motivation is that,analyzing visual information in three-dimensions has enormous benefits,promising more reliable and detailed analytics than in two dimensions(2D).However,breakthroughs in 3D computer vision are marked by advances in depth estimation techniques.One significant approach to depth estimation in 3D computer vision is stereo vision which operates on the basis of visual disparity between two views(stereo images)of a scene.In order to obtain accurate visual disparity cues,it is necessary to find the image coordinates of corresponding points in the two views of the scene.The solution to this correspondence problem is one of the challenging problems in computer stereo vision.This dissertation presents a comprehensive study of depth estimation in computer stereo vision.The work first provides a concise and unambiguous description of two-view geometry that is required to understand the embedded geometric problem in computer stereo vision.We further provide theoretical investigations to show the impact of stereo camera baseline on the computational complexity and accuracy in depth estimation task.Furthermore,the thesis provides a detailed study of stereo correspondence problem(stereo matching)which is the heart of computer stereo vision approach to depth estimation.Crucially,we contribute robust techniques in both traditional and deep learning stereo matching frameworks to alleviate some of the crucial hurdles in stereo matching task.We summarize the distinctive flavors and innovative ideas in these contributions as follows:Firstly,we propose a bidirectional stereo matching method with an adaptive edge-aware cost filter that improves the computational efficiency and accuracy of disparity(depth)prediction.In particular,as opposed to existing approach,the proposed method provides a matching framework that allows simultaneous computation of left and right matching cost volumes.Additionally,we design a more generic framework for adaptive matching cost volume filtering.This scheme is capable achieving a superior trade-off between cost volume smoothing and edge-preservation compared with existing methods.Experimental results show that the approach provides high quality disparity map(depth image)and accelerate processing speed.Secondly,this thesis proposes a robust statistical and probability framework for performing inferences and refinement task on disparity maps in an effort to improve the quality of predicted depth images.Precisely,we design a flexible region-based Markov Random Field(MRF)representation of disparity map.This framework allows for robust analysis and estimations through Bayesian inferences.Moreover,we propose a soft-segmentation constraint and an interpolation scheme for error detection and refinement tasks respectively.A comprehensive experiments and rigorous analysis demonstrate that the proposed method is effective and could significantly improve the quality of predicted noisy disparity maps.Finally,we contribute a robust scale-aware stereo matching network to address two key challenges in deep learning approach to stereo matching task.Recently deep convolutional neural networks(DCNNs),a dominant model in deep learning are proving more effective for stereo matching task than the traditional approaches.However,DCNN models are also being challenged by the inherent ambiguities in stereo correspondence problem caused by inevitable ill-conditioned image textures and the existence of multi-scale objects in real-world scenes.The proposed network focuses on computing multi-scale disparity maps and fusing them to obtain a more reliable and accurate predictions at these challenging situations.To better leverage the complementary advantages of multi-scale disparity maps,we design a robust trainable scale-aware fusion model to integrate confidence maps in the fusion process.The model computes confidence maps by analyzing uncertainties in matching cost volumes and complex disparity relationships among neighboring image pixels.Experiments demonstrate that the proposed stereo matching network achieves superior robustness and generalization performance compared with many recent state-of-the-art methods.In summary,the research work,contributions and findings in this dissertation provide valuable information to boost and advance research in computer stereo vision.