三维点云场景中对象识别技术研究

发布时间：2018-01-02 01:24

本文关键词：三维点云场景中对象识别技术研究　出处：《南京大学》2016年硕士论文　论文类型：学位论文

【摘要】：在复杂的场景中进行物体识别是计算机视觉一个很重要的研究领域。在过去的几十年里,2D物体识别已经被广泛研究,成为一个相对成熟的领域。相比较于2D图像,3D点云可以提供更多几何信息,因此在三维点云中估计物体的姿态比二维图像中估计的姿态更加准确。同时,廉价3D获取设备(如：Kinect)的快速发展也使得点云数据更容易获取。基于这些优势,三维点云物体识别逐渐成为一个研究热点,在机器人导航、虚拟现实、人机交互、场景理解、激光遥感测量等众多领域都有广泛应用。现有的3D点云物体识别方法主要分为两大类：基于局部特征的方法和基于全局特征的方法。基于局部特征的方法是在关键点上提取局部特征描述符进行匹配,所以对场景的遮挡更加鲁棒,但是特征点的错配问题非常严重。基于全局特征的方法在识别过程中将物体作为一个整体来处理,所以速度非常快,但是物体的姿态估计却成为另一个技术难点。针对上述问题,本文将图匹配应用于特征点匹配问题,并充分发挥局部特征和全局特征方法的优势,基于点云库(Point Cloud Library,PCL)实现了上述问题的解决方案,主要工作包括以下几个方面：1.提出了一种基于图匹配的局部特征对象点云识别技术。基于局部特征方法关键的一步就是特征点匹配中如何去筛选出正确的匹配对。几何一致性算法是经典的解决特征点匹配的方法,但是它每次迭代只考虑一组匹配对,这样只能达到局部最优的结果。我们将图匹配应用到点云识别中来,并且采用一种快速地近似图匹配策略,从全局的角度来解决点云识别中的特征点匹配问题,并且用实验验证取得了不错的效果。2.提出了一种全局特征和局部特征融合的点云识别技术。以机器人场景理解为导向的点云对象识别对实时性和定位的准确性要求都比较高,所以我们综合了全局特征和局部特征识别的特点,使用全局特征进行对象的快速识别,并用局部特征进行目标定位。最后对实验结果进行了分析,实时性和准确性都得到满足。3.开发了点云识别原型系统。以对象识别技术为主体,结合一系列点云处理算法,基于Qt开发了点云识别原型系统。系统实现了点云数据和特征直方图的可视化、预处理、模型训练、特征计算、分割、对象识别等功能。
[Abstract]:In a complex scene object recognition is an important research field of computer vision. In the past few decades, 2D object recognition has been studied extensively, and has become a relatively mature field. Compared with 2D images, 3D point cloud can provide more geometric information, the 3D point cloud object estimation the attitude estimation is more accurate than the two-dimensional image attitude. At the same time, cheap 3D acquisition equipment (such as: Kinect) the rapid development of the point cloud data easier to access. Based on these advantages, 3D point cloud object recognition has become a hotspot in robot navigation, virtual reality, human-computer interaction, scene understanding in many areas, the laser remote sensing measurements are widely used. The existing methods of object recognition 3D point cloud is mainly divided into two categories: methods based on local features and method based on global features based on local features. The method is the key point in the extraction of local feature descriptor matching, so the scene is more robust to occlusion, but the problem of mismatch points is very serious. Methods based on global features as a whole in the process of identifying objects to deal with, so the speed is very fast, but the object pose estimation has become another technology the difficulty. To solve the above problems, this paper will be applied to the graph matching feature point matching problem, and give full play to local and global feature method advantage, point cloud based on Library (Point Cloud Library, PCL) to achieve solutions to the above problems, the main work includes the following aspects: 1. propose a local feature recognition technology the object point cloud based on graph matching algorithm based on local feature method. The key step is feature matching to filter out the correct matching. Geometric consistency algorithm is Method to solve the feature point matching code, but it only considers each iteration a set of matching, it can only achieve the local optimal results. We will identify the point cloud in the application of graph matching, and by using a fast approximate graph matching strategy, from a global perspective to solve the matching problem of feature recognition in point cloud and the experiment is used to validate the achieved good results of the proposed.2. point cloud recognition technology is a fusion of global and local features. The robot scene understanding accuracy of point cloud object recognition based on real-time and positioning are high, so we combined the global feature and local feature recognition. Fast recognition using global features of an object, and the target localization with local features. Finally, the experimental results were analyzed, real-time and accuracy have been developed to meet the.3. point cloud The prototype system is identified. Taking object recognition technology as the main body and combining with a series of point cloud processing algorithms, a point cloud recognition prototype system is developed based on Qt. The system realizes the visualization of point cloud data and feature histogram, preprocessing, model training, feature computation, segmentation, object recognition and so on.

【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】