【作者】 王超；

【导师】 杨龙；

【作者基本信息】 西北农林科技大学 ， 农业硕士（专业学位）， 2022， 硕士

【摘要】 真实场景的高质量三维扫描重建和绘制是增强现实、混合现实和机器人应用的重要基础。消费级RGB-D相机三维扫描以便捷、低成本突破了专业三维扫描设备感知范围受限以及对扫描环境要求严格的约束,满足普通用户进行场景信息获取的需求。2011年微软推出的Kinect Fusion方法是消费级深度相机三维扫描重建领域的一项开创性工作,该方法使用一台Kinect深度相机实现了围绕物体级目标进行实时扫描和重建。虽然消费级RGB-D相机的移动式扫描为重建目标提供了新手段,但场景级三维目标扫描重建通常需要融合所有采集深度图、面临数据量大、相机匹配误差导致重建结果漂移、以及耗时高的不足。同时,现有三维扫描重建技术假设相机慢速平稳运动,对相机运动有较高要求,使用条件苛刻。针对上述问题,本文提出一种使用消费级深度相机的场景级三维目标扫描重建方法。本文主要研究内容及贡献概述如下:(1)将稀疏深度图序列融合思想推广至场景级三维目标扫描重建中。通过分析相机运动姿势,对捕获的原始深度图序列进行采样,筛选关键深度图同时剔除抖动帧并减少冗余帧,极大地减少了重建模型的误差累积及漂移现象。(2)研究面向目标几何特征的点云片段配准方法。在关键深度图序列上构造滑动窗口,在每个窗口内执行关键帧匹配融合从而生成表面片段;利用特征描述符匹配求解表面片段间的配准,通过局部多片段间的连续迭代配准优化各表面片段的全局相机位姿和各片段内关键深度图的相机位姿;最终只使用关键深度图融合目标,生成场景三维表面。(3)提高扫描融合方法的时间和空间效率。筛选关键深度图的操作去除了大量的冗余帧,解决对原始深度图序列的有效压缩问题;计算表面片段的几何特征优化了点云片段间的粗略配准,保持高精度配准的同时显著提升了效率。在消费级深度相机采集的深度图序列与Augmented ICL-NUIM、Scene NN和Stan-ford 3D Scene三类公开数据集上进行测试,将关键序列融合与原始序列融合方法比较,实验结果表明,所提方法可将配准过程的均方根误差降低16%～28%,使用8%～54%的数据量即可完成关键序列融合,运行时间平均缩短56%,同时增强了扫描过程的有效性和鲁棒性,显著地提高了扫描场景的重建质量。更多还原

【Abstract】 High-quality 3D reconstruction and rendering of real scenes is an important foundation for AR,MR and robot applications.The consumer RGB-D camera 3D scanning breaks through the limited sensing range of the professional 3D scanning devices and the strict constraint of scanning setting with convenience and low cost,so as to meet the needs of ordinary users for scene information acquisition.The Kinect Fusion method launched by Microsoft in 2011 is a pioneering work of 3D reconstruction,which uses a Kinect to achieve real-time scanning and reconstruction around object-level targets.Although the consumer RGB-D cameras provide a new means for target reconstruction,3D scanning and reconstruction of scene-level targets usually requires the fusion of all ac-quired depth images.It often confronts several bottlenecks including a large amount of redun-dant data,feature drifting as well as time-consuming.At the same time,the basic assumption behind 3D reconstruction methods is that the camera moves slowly and smoothly,it has high requirements on camera movement and harsh conditions of use.To solve these problems,a scene-level targets reconstruction method using a consumer depth camera is proposed.The main contents and contributions include:(1)The fusion of sparse sequence is extended to scene-level targets reconstruction.Based on the analysis of camera trajectory it constructs the key subset for depth image sequence,which eliminates jittery frames and reduces redundant frames via sampling the depth image sequence.Our method greatly reduces the error accumulation and drift.(2)Propose the point cloud fragment registration method based on geometric features.The key subset is divided into a set of successive sliding windows,surface fragment is gener-ated by key frames matching and fusion within each window.To optimize the camera motion trajectory,geometric features are introduced to the process of the continuous iterative regis-tration between multiple fragments.Finally,fusing the key subset could generate the targeted surface.(3)Our method improves the time and space efficiency of scan fusion.Sampling the depth image sequence removes a large amount of redundant data and solves the problem of efficient compression of the original depth images sequence;the geometric features of fragments are calculated to optimize the coarse registration between fragments,maintaining high-precision registration while significantly improving the efficiency.The scanning tests and the comparison experiments are conducted on depth image se-quence captured by a consumer depth camera and three public datasets,Augmented ICL-NUIM,Scene NN and Stanford 3D Scene.The results show that the proposed method can reduce the registration RMSE by 16%～28%,and use only 8%～54%data to complete the key sequence fusion.The running time is shortened by about 56%.In addition,it enhances the effectiveness as well as robustness of 3D scanning and improves the reconstruction quality significantly.更多还原