基于深度相机的场景物体定位与抓取研究

发布时间：2018-01-10 04:00

本文关键词：基于深度相机的场景物体定位与抓取研究　出处：《南京大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,深度摄像头因其功能强大、成本相对较低、部署方便,在室内场景建模、机器人自动定位导航等领域得到广泛使用。但是,在利用深度摄像机的深度视觉数据来驱动机械臂进行物体抓取方面,由于对精度和实时性有较高要求,现有方法较难满足要求。本文对这一问题进行了系统研究,论文选题具有较强的理论意义和实用价值。本文提出了一种基于3D点云配准、深度摄像头驱动的机械臂自动物体抓取方法。具体来讲,本文结合RGB-D数据的相关特点,对基于视觉驱动的机械臂定位与抓取问题进行了综述,梳理出当前机械臂领域的研究与应用现况。之后,本文给出了一种新的深度数据驱动的机械臂抓取算法框架,并将其分为区域分割、物体分类、点云配准以及机械臂驱动四个算法模块,具体工作包括:1.预处理与3D抓取区域的标注:使用3D扫描仪对常见桌面级物体进行建模,并在3D点云模型上标注出物体的可抓取部位;之后采集物体的二维图片,以训练分类器;2.结合深度与颜色信息的物体筛选:利用深度摄像头采集的深度信息对二维图片进行区域分割,再利用预处理时生成的物体分类器进行分割后的区域进行分类,得到物体的二维候选区域;3.点云配准:利用深度信息和摄像机的内参生成点云数据,进而利用生成的点云数据和扫描物体的点云数据进行点云配准,获取空间的转移矩阵,最终通过转移矩阵将物体的可抓取部分映射回深度相机的三维空间;4.视觉驱动的机械臂抓取:利用棋盘标定来估算外参后,将相机空间预测的区域转换到机械臂空间,从而驱动机械臂实施抓取任务。本实验采集了五类常见的桌面级物体(瓶子、茶杯、易拉罐、茶壶和盒子),分别对其建模并利用提出的算法进行了测试,并利用杰卡德相似度作为评判标准。实验结果表明,本文所提出的方法取得了良好的效果。
[Abstract]:In recent years, depth camera has been widely used in many fields such as indoor scene modeling, robot automatic positioning and navigation because of its powerful function, relatively low cost and convenient deployment. In the aspect of using depth vision data of depth camera to drive the robot arm to grab objects, there are high requirements for accuracy and real-time. The existing methods are difficult to meet the requirements. This paper makes a systematic study of this problem, the topic of this paper has strong theoretical significance and practical value. In this paper, a method based on 3D point cloud registration is proposed. Depth camera driven robotic arm automatic object capture method. Specifically, combining with the characteristics of RGB-D data, this paper summarizes the problem of robot arm location and capture based on visual drive. After combing out the current research and application status in the field of manipulator, this paper presents a new depth data-driven arm grab algorithm framework, and divides it into region segmentation and object classification. Point cloud registration and robot arm drive four algorithm modules, the specific work includes: 1. Preprocessing and 3D capture region tagging: using 3D scanner to model common desktop objects. At the same time, the grabable part of the object is marked out on the 3D point cloud model. Secondly, the two-dimensional images of the objects are collected to train the classifier. 2. Object selection based on depth and color information: using depth information collected by depth camera to segment two-dimensional images. Then, the object classifier generated by preprocessing is used to classify the segmented regions, and the candidate regions of the objects are obtained. 3. Point cloud registration: point cloud data are generated by using depth information and camera parameters, and then point cloud registration is carried out by using the generated point cloud data and the point cloud data of scanned objects, and the spatial transfer matrix is obtained. Finally, the grabable part of the object is mapped back to the 3D space of the depth camera through the transfer matrix. 4. Visual driven manipulator grab: after the outer parameters are estimated by the checkerboard calibration, the area of camera space prediction is converted to the manipulator space. In this experiment, five kinds of common desktop objects (bottles, teacups, cans, teapots and boxes) were collected, and their models were modeled and tested by using the proposed algorithm. The experimental results show that the proposed method has good results.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】