SIA OpenIR  > 光电信息技术研究室
基于三维点云的目标识别与定位方法研究
Alternative TitleResearches on Target Recognition and Localization Method Based on Three-dimensional Point Cloud
鲁荣荣1,2
Department光电信息技术研究室
Thesis Advisor朱枫 ; 吴清潇
Keyword目标识别 点云分割 点对特征 三维局部特征 几何约束
Pages112页
Degree Discipline模式识别与智能系统
Degree Name博士
2019-05-14
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract随着《中国制造2025》制造强国战略的提出,以“个性化定制、柔性化生产”为理念的小批量生产模式已成必然趋势。机器人作为实现这一转变的关键要素,提升机器人的环境感知能力迫在眉睫。其中,抓取物体是机器人与环境自主交互过程中十分基础但很重要的环节,而机器人完成抓取操作的前提是其视觉感知系统能够对目标进行识别定位。以模板匹配为代表的二维目标识别定位方法虽然在传统的大批量、结构化生产模式中发挥了关键作用,但它们并不能很好地处理非结构化环境下的目标识别定位问题。三维目标识别定位方法以其更加灵活、更加鲁棒、更加精确的识别定位性能成为提升机器人环境感知能力的首选。开展三维目标识别定位方法的研究具有重要的理论意义与实际应用价值。三维目标识别在过去的三十年里已取得较大的理论突破,但其中仍有一些环节有待进一步改进与完善。本文以三维点云为主要研究对象,结合实际工程应用需求,针对三种类型的目标进行识别定位方法的研究,分别是几何特征单一的平面型物体;几何特征较丰富的物体以及几何特征丰富的物体,并据此凝练出以下四个研究内容:(1) 基于有序点云分割的平面型物体识别定位方法。平面型物体可以看成将某种二维形状垂直拉伸一定高度后形成的立体构型。这类物体在工业生产中十分常见,且多为需要进一步加工的原材料,例如铺设地面的板砖、经切削后需要打磨的钢板等。在自动化生产加工中,视觉系统需要根据传感器采集的数据对它们进行识别定位,以辅助机械手完成抓取操作。对于多个平面型物体混叠摆放的场景,基于局部特征匹配的识别定位方法在处理结构简单、几何特征单一的平面型物体时,容易出现误匹配而导致识别失败;基于模板匹配的方法因为效率不高,抗遮挡能力不够,识别定位效果也不理想。针对这一问题,本课题提出一种基于有序点云分割的平面型物体识别定位方法,以实现对这类物体的快速识别定位。(2) 基于增强型点对特征的三维目标识别定位方法。点对特征特指由两个点及其法向量构建的四元属性组。由于其组成要素少,生成方式快捷简单,使其具有高效、灵活的曲面表征能力,可以用来描述几何特征并不丰富的三维点云模型。本课题在基于点对特征投票的三维目标识别算法基础上,针对其中因点对特征存在二义性导致算法识别率不高以及模型点对特征哈希表中存在大量冗余点对导致内存浪费的问题,提出一种在视点可见性约束下,基于增强型点对特征的三维目标识别定位方法,以提高算法的计算效率和识别准确率。(3) 基于三维位置信息的局部特征描述方法。三维局部特征描述是3D视觉领域的重点研究方向之一,它是很多高级任务的基础,例如三维目标建模、三维目标检索以及三维目标识别等。评价一个三维局部特征描述子的综合性能通常从以下五个维度出发:区分性、抗干扰性、紧凑性、计算高效性以及可扩展性。其中,区分性主要考察特征的表征能力;抗干扰性主要考察特征对噪声、遮挡等因素的抵抗能力;紧凑性主要考察特征的维度是否足够低,以便于存储和特征匹配;计算高效性主要考察特征的生成方式是否快捷;可扩展性主要考察特征对于多模态信息能否灵活兼容。在过去的二十多年里,诞生了许多性能优异的三维局部特征描述方法。但基于上述五个维度对它们进行考察,发现每种三维局部特征描述方法至少存在一个方面的短板。针对这一问题,本课题在总结前人优秀工作的基础上,提出一种基于三维位置信息的局部特征描述方法,试图在这五个维度上取得很好的平衡。(4) 基于几何约束与投票的位姿估计方法。在基于三维局部特征匹配的目标识别方法中,估计目标的6D位姿需要依赖正确的匹配关系。性能优异的三维局部特征有助于在特征匹配环节建立更多正确的匹配对,但无法避免引入错误的匹配。引起误匹配的原因有多种,例如场景中存在重复结构、场景点云带有噪声、场景中其他物体的干扰、目标特征显著的地方被遮挡、目标自身的相似性干扰以及三维局部特征的表征能力不够等。错误的匹配会影响位姿估计的准确率和精度,现有的误匹配剔除算法在准确率和效率方面还存在优化改进的空间。本课题针对现有方法的不足,提出一种基于几何约束与投票的位姿估计方法,试图将位姿估计与误匹配剔除联合优化,从而提升位姿估计的准确率和精度。该方法不止适用于三维目标识别,在点云配准、三维目标建模等任务中也能发挥作用,具有一定的通用性。
Other AbstractWith the proposal of the "Made in China 2025" manufacturing strong country strategy, the small batch production mode with the concept of "individualized customization and flexible production" has become an inevitable trend. As a key element to achieve this transformation, robots are urgently needed to enhance the environmental awareness of robots. Grasping an object is the most basic but also the most important part of an industrial robot's autonomous task. Identifying and locating the target is the prerequisite for the robot to complete the grab operation. Although the two-dimensional target recognition and localization method represented by template matching plays a key role in the traditional high-volume, structured production mode, they cannot deal well with the target recognition and localization problem in the unstructured environment. The 3D target recognition and localization method is the first choice for improving the perception of the robot environment with its more flexible, more robust and more accurate recognition and localization performance. The research on the three-dimensional target recognition and location method has important theoretical significance and practical application value. Three-dimensional target recognition has made great theoretical breakthroughs in the past three decades, but there are still some links that need further improvement and improvement. In this paper, the three-dimensional point cloud is the main research object. Combined with the actual engineering application requirements, the research on the recognition and positioning methods for three types of targets is a planar object with a single geometric feature; a rich geometric object and a rich geometric feature. Objects, and based on this, condensed the following four research contents: (1) Research on planar object recognition and localization method based on ordered point cloud segmentation. A planar object can be seen as a three-dimensional configuration formed by stretching a certain two-dimensional shape vertically to a certain height. Such objects are very common in industrial production, and are mostly raw materials that require further processing, such as floor tiles that are laid on the ground, and steel plates that need to be polished after cutting. In automated production processing, the vision system needs to identify and locate the sensors based on the data collected by the sensors to assist the robot in completing the grab operation. For scenes in which multiple planar objects are placed in an overlapping manner, the localization method based on local feature matching is easy to mismatch and cause recognition failure when dealing with planar objects with simple structure and single geometric features. The method based on template matching is because the efficiency is not high, the anti-blocking ability is not enough, and the recognition and localization effect is not ideal. Aiming at this problem, this paper proposes a planar object recognition and localization method based on ordered point cloud segmentation to realize fast recognition and location of such objects. (2) Research on 3D target recognition and localization based on enhanced point pair features. A point pair feature specifically refers to a quaternion of attribute groups constructed from two points and their normal vectors. Because of its few components, the generation method is quick and simple, which makes it have efficient and flexible surface representation ability, which can be used to describe the 3D point cloud model with not rich geometric features. Based on the 3D target recognition algorithm based on point-to-feature voting, the problem is that the recognition rate of the algorithm is not high due to the ambiguity of the point-to-feature feature and the memory point is invalid due to the existence of a large number of redundant pairs in the feature point hash table. The problem is to propose a three-dimensional target recognition and localization method based on enhanced point-pair features under the constraints of viewpoint visibility to improve the computational efficiency and recognition accuracy of the algorithm. (3) Research on local feature description method based on three-dimensional position information. Three-dimensional local feature description is one of the key research directions in the field of 3D vision. It is the basis of many advanced tasks, such as 3D target modeling, 3D target retrieval and 3D target recognition. Evaluating the comprehensive performance of a three-dimensional local feature descriptor usually starts from the following five dimensions: distinguishing, anti-interference, compactness, computational efficiency, and scalability. Among them, the distinguishing nature mainly investigates the characterizing ability of the feature; the anti-interference mainly investigates the resistance of the feature to noise, occlusion and other factors; the compactness mainly checks whether the dimension of the feature is low enough to facilitate storage and feature matching; Whether the feature is generated quickly or not; the scalability mainly examines whether the feature is flexible and compatible with multimodal information. In the past twenty years, many excellent three-dimensional local feature description methods have been born. However, based on the above five dimensions, they are found to have at least one shortcoming in each aspect. In response to this problem, this topic proposes a local feature description method based on three-dimensional position information on the basis of summarizing the excellent work of the predecessors, trying to achieve a good balance in these five dimensions. (4) Research on pose estimation method based on geometric constraints and voting. In the target recognition method based on three-dimensional local feature matching, estimating the 6D pose of the target needs to depend on the correct matching relationship. The excellent three-dimensional local features help to establish more correct matching pairs in the feature matching, but cannot avoid introducing wrong matching. There are several reasons for mismatching, such as the existence of repetitive structures in the scene, the noise of the scene point cloud, the interference of other objects in the scene, the occlusion of the target features, the similarity of the target itself, and the characterization of the three-dimensional local features. Not enough capacity. The wrong matching will affect the accuracy and accuracy of the pose estimation. The existing mismatch culling algorithm still has room for optimization and improvement in terms of accuracy and efficiency. Aiming at the shortcomings of the existing methods, this paper proposes a pose estimation method based on geometric constraints and voting, trying to optimize the pose estimation and mismatch matching, so as to improve the accuracy and accuracy of pose estimation. This method is not only suitable for 3D target recognition, but also plays a role in tasks such as point cloud registration and 3D target modeling, and has certain versatility.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/25149
Collection光电信息技术研究室
Affiliation1.中国科学院沈阳自动化研究所
2.中国科学院大学
Recommended Citation
GB/T 7714
鲁荣荣. 基于三维点云的目标识别与定位方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
基于三维点云的目标识别与定位方法研究.p(6810KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[鲁荣荣]'s Articles
Baidu academic
Similar articles in Baidu academic
[鲁荣荣]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[鲁荣荣]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.