CASIA OpenIR
视觉三维地图变化检测和更新算法研究
和颖
2021-05-25
页数59
学位类型硕士
中文摘要

基于图像的室内外场景三维重建和定位导航技术一直是计算机视觉领域的热点研究领域之一。现有的许多实际应用,如数字城市、智能机器人、无人机、自动驾驶、增强/虚拟现实等,都对场景三维地图的准确性和时效性提出了很高要求。在现有的研究中,大量的研究工作集中于高效精准的三维地图构建方法,但对于如何保持三维地图的时效性,即如何有效地发现地图中的变化并及时更新,还没有系统性的解决方案。针对这一瓶颈问题,本文针对视觉三维地图变化检测和更新算法开展了系统性研究,并在真实场景中进行了实验验证,主要工作如下:

针对视觉定位领域常用的稀疏三维地图,本文提出了一种融合图像特征点匹配和点云几何一致性度量的稀疏地图变化检测和更新方法。该方法首先将新图像批量配准到现有的三维地图中,并与空间位置临近的旧图像进行局部引导的特征匹配,以此度量稀疏地图中3D点对应的图像特征匹配一致性;之后,将三维地图进行栅格化划分,并整体度量每个栅格内稀疏3D点的特征匹配一致性和几何一致性;最后通过马尔可夫能量优化对栅格变化与否进行标记,并将标记为变化的栅格可见图像替换为新图像,完成稀疏三维地图变换检测和更新。实验结果表明,该方法可以有效检测室内外场景稀疏三维地图中的变化区域,相比于数据驱动的端到端变化检测方法具有更好的场景适应能力。

除稀疏三维地图外,针对地理信息、机器人等领域常用的稠密三维地图,本文进一步提出了一种基于虚拟渲染和深度稠密描述子相似性度量的稠密三维地图变化检测和更新方法。该方法首先将新图像批量配准到现有的三维地图中,并在新图像位置通过OpenGL渲染生成三维地图对应的虚拟图像;然后通过逐像素的稠密描述子匹配衡量新图像和虚拟图像之间的像素级一致性;之后通过反投影和空间几何一致性融合,度量每个空间三角面片的变化程度,并通过全局优化方法标记变化区域;最后将标记为变化的空间三角面片可见图像更新为新图像,实现稠密三维地图的更新。实验结果表明,该方法可以有效利用场景语义分割网络提供的深度稠密描述子发现场景中各类细节变化。

英文摘要

Image-based 3D reconstruction of indoor and outdoor scenes and Localization and navigation technology has always been one of the hot research fields in computer vision. Many existing practical applications, such as digital cities, intelligent robots, unmanned aerial vehicles, autonomous driving, augmented/virtual reality, etc., put forward high requirements for the accuracy and timeliness of scene 3D map. In the existing research, a large number of research works focus on efficient and accurate 3D map construction methods, but there is no systematic solution for how to maintain the timeliness of 3D maps, that is, how to effectively detect changes in the map and update them in time. To solve this bottleneck problem, this paper carried out a systematic study on the change detection and update algorithm of 3D visual map, and carried out experimental verification in real scenes. The main work is as follows:

Aiming at the commonly used sparse 3D maps in the field of visual localization, this paper proposes a sparse map change detection and update method that integrates feature matching and point cloud geometric consistency measurement. Firstly, the new images are registered to the existing 3D maps in batches, and the new images are matched with the old images which are close to each other in space. In this way, the texture consistency of the 3D points in the sparse map is measured. Then, the 3D map was divided to grids, and the texture consistency and geometric consistency of the sparse point clouds in each grid were measured as a whole. Finally, Markov energy optimization is used to mark whether the grid changes or not, and the visible images of grids marked as changes are replaced with new images to complete the detection and update of map transformation. Experimental results show that the proposed method can effectively detect the changing areas of the 3D sparse map in indoor and outdoor scenes, and has better scene adaptability compared with the data-driven end-to-end change detection method.

In addition to sparse 3D maps, this paper further proposes a change detection and update method based on virtual rendering and dense feature matching for dense 3D maps, which are commonly used in geographic information and robot fields. Firstly, the new images are registered into the existing 3D maps in batches. After that, the virtual image corresponding to the 3D map is generated by OpenGL rendering with the new image position. Then the pixel-level consistency between the new image and the virtual image is measured by the per-pixel matching of the dense descriptor. Finally, the changed degree of each facet was measured by the fusion of back-projection and spatial geometric consistency, and the changed region was marked by the global optimization method. Finally, the visible image of facets marked as change is replaced by new images to update the dense map. Experimental results show that the proposed method can effectively use the deep dense descriptors provided by the scene semantic segmentation network, to detect the changes of various details in the scene.

关键词稀疏三维地图 稠密三维地图 变化检测 地图更新
语种中文
七大方向——子方向分类三维视觉
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/44993
专题中国科学院自动化研究所
模式识别国家重点实验室
推荐引用方式
GB/T 7714
和颖. 视觉三维地图变化检测和更新算法研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2021.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
ucasthesis_master.pd(18320KB)学位论文 开放获取CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[和颖]的文章
百度学术
百度学术中相似的文章
[和颖]的文章
必应学术
必应学术中相似的文章
[和颖]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。