SIA OpenIR  > 光电信息技术研究室
面向机器人航天员的三维目标识别定位方法研究
Alternative TitleResearch on 3D Object Recognition and Location Method for Robot Astronaut
潘旺
Department光电信息技术研究室
Thesis Advisor朱枫
Keyword识别定位 位姿测量 三维目标 模板匹配 多传感器
Pages99页
Degree Discipline模式识别与智能系统
Degree Name博士
2018-11-27
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract

三维目标识别定位是机器人视觉领域的重要研究内容,其中三维目标识别是指在传感器获取的数据中辨认出目标物体;定位是指确定出目标物体在三维空间里的位置和姿态。物体特征是进行目标识别和定位不可或缺的要素,包括颜色、纹理、角点、形状等。无纹理物体在工业、航天、日常生活等领域很常见,逐渐成为热门研究对象。因其纹理特征匮乏,所以形状作为一种通用特征,目前被广泛使用。基于形状的模板匹配方法在二维目标识别与定位方面已取得了许多重要研究成果,且部分成果在工业领域的特定场景下已成熟应用。但因为三维目标带来的复杂性,此类方法应用于三维领域还存在诸多未解决的理论问题,它依然是计算机视觉研究领域的主要方向之一。本文面向机器人航天员应用,对三维目标识别定位方法开展相关研究,具有非常重要的理论意义和实际工程应用价值。三维目标识别定位要考虑的因素很多,如精度与速度的折中、特征的可区分度、大型复杂场景、三维物体的视图退化问题等。本文以实际的工程应用需求为背景,在不同条件下对识别定位方法开展了研究:(1) 基于三维CAD模板匹配算法的参数优化方法。基于特征描述子的目标识别定位方法在某些领域已广泛应用,但其无法处理表面光滑无纹理的物体,因为在成像中无法提取到可用的特征描述子。目前最先进的基于三维CAD模板匹配算法,将边缘点梯度方向作为特征,融合了尺度空间理论和相似度评价函数,具有计算速度快,对光照、噪声和遮挡鲁棒性强等优点。该方法在特定的工业应用场景中取得了成功,但在大场景或复杂背景下,该方法存在错误率高,稳定性差,搜索效率低等问题。为了改善这些问题,研究算法中各参数对识别定位鲁棒性、精度和速度的影响,对参数优化进行研究对算法的实际工程应用具有重要意义和参考价值。(2) 基于多表观特征的三维目标识别定位方法。表面光滑、无纹理的三维物体的成像特征较少,识别难度较大。目前常用的基于梯度方向的形状模板匹配方法适用于各种不同形状的物体。但由于梯度方向与物体的形状结构有关,可区分性并不强,不足以充分表征目标物体。尤其是在大场景、复杂混乱背景下,极易出现误匹配,将背景中的非目标区域误判成目标物体,导致识别定位结果错误,鲁棒性差。因此可以通过提高物体特征的丰富性和表征的可辨识性,增强识别定位的准确性,提高正确率。所以,研究目标物体多种图像特征的融合方法以避免错误识别定位。(3) 基于深度图预分割的三维目标识别定位方法。目前三维目标的识别定位存在搜索空间巨大,搜索效率低等问题,部分改进方法为了提高速度,牺牲定位精度,将图像特征进行泛化或卷积处理,无法实现三维目标高精度位姿测量。深度相机的分辨率较低,采集到的数据精度有限且噪声较大,且视觉特征没有二维图像直观。而CCD相机采集的二维图像数据分辨率和精度均优于深度相机,但用于三维目标位姿测量时存在搜索空间大,效率低等问题。针对精度和速度难以两全的问题,研究基于多传感器的三维目标位姿测量方法,以实现三维目标快速高精度位姿测量。(4) 基于点云提取的矩形目标高精度位姿测量方法。针对基于单目视觉的形状模板匹配方法用于三维目标位姿测量时,退化视图易出现误匹配导致位姿测量失败的问题,提出了一种基于多传感器的三维目标位姿测量方法。充分发挥深度相机和高分辨率工业相机各自的优势,两种成像技术形成互补,提高了测量的鲁棒性和效率。首先利用物体与其固定平面之间的关系,在点云中粗略定位出目标区域,然后通过预先标定信息将目标区域转换到灰度图像空间。在灰度图像中,利用LSD直线提取算法外加特征约束,筛选出四条目标直线,最终利用P4P算法求解出目标的位姿。实验验证了该算法的有效性,改善了退化视图下的错误测量问题,且测量效率远优于经典的模板匹配方法。本文的研究工作面向机器人航天员,对已知模型的无纹理三维目标识别与位姿测量算法进行了深入探讨和研究。在上述理论研究的基础上,应用前面提出的算法实现了对目标物体的快速高精度位姿测量,验证了理论方法的可行性,有助于推动此类方法在工程实践中的应用。

Other Abstract

Three-dimensional object recognition and localization are important research aspects of vision systems in robotic applications. Three-dimensional object recognition refers to identifying the target object in the data acquired by the sensor. Localization refers to determining the position and orientation of the target object in three-dimensional space. Features are indispensable elements for object recognition and localization, including color, texture, corner, shape and so on. Textureless objects, which are common in industries, aerospace and daily life, have gradually become popular research objects. Because of its lack of texture features, shape is widely used as a visual feature. Shape-based template matching methods have achieved many important research results in two-dimensional target recognition and localization, and some of the results have been applied in specific scenes in the industrial field. However, due to the complexity of the three-dimensional object, there are still many unsolved theoretical problems when such methods are used in the three-dimensional field. It remains one of the main directions in the field of computer vision research. This paper takes the robot astronaut as the application background, and carries out related research on the three-dimensional object recognition and localization method, which has very important theoretical significance and practical engineering application value. There are many factors to consider in three-dimensional object recognition and localization, such as the compromise between accuracy and efficiency, the distinguishability of features, large-scale and complex scene and the problem of view degradation of three-dimensional objects. Based on the actual engineering application requirements, this paper studies the recognition and localization methods under different conditions: (1) Research on parameter optimization of template matching algorithm based on three-dimensional CAD. Object recognition and localization methods based on feature descriptors have been widely used in some fields, but they cannot deal with smooth and textureless objects because available feature descriptors of objects cannot be extracted in the image. At present, the most advanced 3D CAD template matching algorithm takes the gradient direction as the feature, and combines the scale space theory and the morphological graph method based on the similarity evaluation function. It has the advantages of fast speed, strong robustness to illumination, noise and occlusion. This method has been successfully used in specific industrial application scenarios, but in large scenes or complex backgrounds, the method has problems of high recognition error rate, poor robustness and slow speed. In order to improve these problems, the effects of various parameters in the identification and location algorithm on robustness, accuracy and speed are studied. Research on parameter optimization is of great significance and reference value for the practical engineering application of the algorithm. (2) Three-dimensional target recognition and localization method based on multiple apparent features. The smooth and textureless three-dimensional object has fewer imaging features and is more difficult to identify. The current shape template matching method based on gradient direction is applicable to objects of various shapes. However, since the gradient direction is related to the shape and structure of the object, the distinguishability is not strong enough to fully characterize the target object. Especially in the context of large scenes with complex chaotic background, it is easy to mismatch, that is, the non-target area in the background may be misjudged into the target object, resulting in wrong recognition result and poor robustness. Therefore, by increasing the variety of the feature of the object and the identifiability of the representation, the robustness of the recognition and localization is enhanced, and the correct rate is improved. Therefore, the fusion method of multiple image features of the target object is studied to avoid the mistake in identification and localization. (3) Three-dimensional target recognition and localization method combining depth map and gray image. At present, the recognition and localization of 3D targets has many problems such as difficulties in huge search space and low search efficiency. Some improved methods can improve the speed by sacrificing the positioning accuracy, and generalize or convolve the image features, which can not achieve high-precision pose measurement of 3D targets. The resolution of the depth camera is low, the acquired data is limited in precision and has much noise, and the visual features are not as intuitive as the two-dimensional image. The resolution and accuracy of the 2D image data acquired by the CCD camera are better than those of the depth camera, but there are problems such as large search space and low efficiency when used for 3D target pose measurement. Aiming at the problem that high recognition rate and high efficiency can hardly be met simultaneously, the multi-sensor based three-dimensional target pose measurement method is studied to realize the fast and high-precision pose measurement of the three-dimensional target. (4) High-precision pose measurement method based on point clouds and grayscale images. When template matching methods based on monocular vision are used for the pose measurement of a three-dimensional object, degenerated views are very prone to false match, resulting in the failure of pose measurement. Aiming at this problem, we propose a method based on multi-sensor to measure the pose of a three-dimensional object. Taking full advantage of the respective advantages of depth cameras and high-resolution industrial cameras, the two imaging technologies complement each other, improving measurement robustness and efficiency. First, the relationship between the object and its fixed plane is used, and the target area is roughly located in the point cloud, and then the target area is converted into the grayscale image space by pre-calibration information. In the gray image, the LSD algorithm is used to extract lines, and by adding constraints, four target lines are selected. Finally, the P4P algorithm is used to solve the pose of the target. The experiment verifies the effectiveness of the algorithm and improves the error measurement problem in the degraded view, and the measurement efficiency is much better than the classical template matching method. The research work of this thesis is aimed at robot astronauts. The 3D textureless object recognition and pose measurement algorithms of known models are deeply discussed and studied. On the basis of the above theoretical research, the fast and high-precision six-dimensional pose measurement of the target object is realized by the algorithm proposed above, which verifies the feasibility of the theoretical method and helps to promote the application of such methods in engineering practice.

Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/23646
Collection光电信息技术研究室
Recommended Citation
GB/T 7714
潘旺. 面向机器人航天员的三维目标识别定位方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2018.
Files in This Item:
File Name/Size DocType Version Access License
面向机器人航天员的三维目标识别定位方法研(3587KB)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
[潘旺]'s Articles
Baidu academic
Similar articles in Baidu academic
[潘旺]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[潘旺]'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.