基于深度自动分层的RGBD序列场景流计算技术研究

发布时间：2018-04-21 04:20

本文选题：场景流 + RGBD图像序列　；参考：《南昌航空大学》2017年硕士论文

【摘要】：3D场景流是空间场景或物体运动的三维运动矢量场,其包含了场景或物体的三维运动与结构信息,在目标运动估计与跟踪、姿态识别、自主避障、路径规划等研究方向具有重要的研究价值,研究成果被广泛用于航空航天、军事、工业、气象、交通以及文物保护等领域。近年来,随着消费级深度传感器的普及,利用RGBD序列估计3D场景流逐渐成为计算机视觉研究领域的热点问题。虽然现有的RGBD序列场景流计算方法能够获取较为准确的估计结果,但是当图像序列包含复杂背景和多个运动目标时,由于现有方法通常采用人工设定深度图像初始分层层数并且得到的初始场景分割图中只包含深度信息,导致并不能完全准确地分割独立运动目标,致使场景流估计效果较差。针对以上问题,本文主要研究基于深度自动分层的RGBD序列3D场景流计算技术,主要研究工作包括:1.首先对3D场景流计算技术的研究背景和现状进行了介绍,然后论述了图像光流与3D场景流的对应关系,并对3D场景流技术的两类主要方法进行了重点分析。2.针对现有3D场景流计算方法在复杂场景下估计效果较差的问题,提出基于光流的深度图像自动分层与分割优化方法。首先设定任意初始分层层数,然后利用K均值聚类计算深度图像初始分割结果,再根据RGB图像序列光流估计结果分别判断、合并相邻层,最终获取深度图像的分层与分割结果。相对于传统的人工分层方法,本文方法不仅能够实现自动分层,而且得到的分割结果是与运动目标相关的,将各个目标独立分层,更利于分层场景流的计算。3.将本文深度图像自动分层和分割优化方法应用在RGBD图像序列分层场景流的计算当中,介绍了其对应的能量泛函,然后详细论述了各个约束项,最后详细介绍了本文方法应用在分层场景流计算中的求解过程。4.采用Middlebury 2003测试图像集、Middlebury 2005测试图像集、SRSF真实场景图像集、RGBD跟踪场景图像集测试本文深度图像自动分层与分割优化在分层场景流计算中的应用效果,同时进一步测试本文方法的有效性。实验结果表明:1)本文分层场景流计算方法实现了自动分层,不需要人工选择分层数以得到最佳分层效果,并且最终得到的场景分割效果更加准确,能够准确的将各运动目标和背景独立分层;2)本文方法计算得到的场景流计算误差更小,最终得到的场景流计算结果也更符合场景的真实3D运动。
[Abstract]:3D scene flow is a three-dimensional motion vector field of space scene or object motion. It contains 3D motion and structure information of scene or object, and can be used in target motion estimation and tracking, attitude recognition, autonomous obstacle avoidance, etc. The research direction of path planning has important research value, and the research results are widely used in aerospace, military, industry, meteorology, traffic and heritage conservation and other fields. In recent years, with the popularity of consumer-level depth sensors, using RGBD sequences to estimate 3D scene flow has gradually become a hot issue in the field of computer vision. Although the existing RGBD sequence scene flow calculation method can obtain more accurate estimation results, but when the image sequence contains complex background and multiple moving targets, Because the existing methods usually use manual to set the initial layer number of depth image and only contain depth information in the initial scene segmentation image, it is not possible to segment the independent moving object completely and accurately, so the effect of scene flow estimation is poor. Aiming at the above problems, this paper mainly studies the RGBD sequence 3D scene flow computing technology based on depth automatic stratification, the main research work includes: 1. Firstly, the research background and present situation of 3D scene flow computing technology are introduced, then the corresponding relationship between image optical flow and 3D scene flow is discussed, and the two main methods of 3D scene flow technology are analyzed emphatically. In order to solve the problem that the existing 3D scene flow estimation methods have poor performance in the estimation of complex scenes, an optical flow-based method for automatic delamination and segmentation optimization of depth images is proposed. First, the number of arbitrary initial layers is set, then the initial segmentation results of depth images are calculated by K-means clustering, and then judged according to the optical flow estimation results of RGB image sequence, the adjacent layers are merged, and the results of delamination and segmentation of depth images are obtained. Compared with the traditional artificial stratification method, this method can not only achieve automatic stratification, but also the segmentation results are related to moving objects. In this paper, the method of automatic stratification and segmentation optimization of depth image is applied to the calculation of hierarchical scene flow of RGBD image sequence, and its corresponding energy functional is introduced, and each constraint item is discussed in detail. Finally, the solution process of the method applied in hierarchical scene flow calculation is introduced in detail. 4. 4. Using Middlebury 2003 test image set and Middlebury 2005 test image set to test the real scene image set RGBD tracking scene image set to test the application effect of the depth image automatic stratification and segmentation optimization in hierarchical scene flow calculation. At the same time, the effectiveness of this method is further tested. The experimental results show that the hierarchical scene flow calculation method in this paper realizes automatic stratification, does not need to manually select the number of layers to get the best result of stratification, and the final result of scene segmentation is more accurate. Each moving object and background can be accurately stratified. (2) the calculation error of scene flow obtained by this method is smaller, and the result of scene flow calculation is more consistent with the real 3D motion of the scene.
【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】