【作者】 李明；

【导师】 谭毅华；

【作者基本信息】 华中科技大学 ， 控制工程， 2020， 硕士

【摘要】 随着人工智能和计算机视觉技术的发展,工业生产的自动化程度也逐步提高。非接触式的自动化测量技术是工业生产中的重要内容,利用计算机视觉技术可以解决自动化测量中的关键问题:目标识别和姿态估计。近年来,基于深度学习的方法不断发展,逐渐取代了人工提取特征的方法。点云数据是三维数据的重要表示方式之一,相比于二维图像,点云数据包含了物体深度等更多信息,它的表达方式也相对简单,可以作为深度学习模型的输入。本论文基于立体视觉技术获取目标三维点云数据,对基于三维点云的目标识别和姿态估计技术展开研究,论文的主要工作如下:首先,对深度相机进行对比选择,介绍双目相机成像原理,搭建双目立体视觉平台并标定后,采集目标场景的点云数据。由于采集到的点云数据存在噪声和干扰,因此要对点云数据进行预处理,本文采用条件滤波、体素滤波和统计滤波对原始点云进行滤波采样,获得了预处理后的点云数据。其次,在Point Net点云分类网络的基础上,提出了MV-Point Net网络模型。该网络以多视角点云数据作为网络输入,采用View-Pooling的方式聚合多视角特征。对模型使用不同的View-Pooling方法和视角数目进行了评估实验,并与Point Net进行对比,在基于Model Net数据集制作的数据上获得了86.5%的识别准确率。使用相机获取到的目标点云进行了在线识别的实验,验证了实际的识别效果。最后,对目标点云与模板点云进行配准得到变换矩阵实现识别目标的姿态估计。提出了深度学习粗配准与ICP精配准结合的方法,该方法精度相对各自独立算法有所提升,与传统的ICP算法相比速度提升了2倍。更多还原

【Abstract】 With the development of artificial intelligence and computer vision technology,industrial production is becoming more and more automatic.Non-contact automatic measurement technology is an important content in industrial production.Computer vision technology can solve the key problems in automatic measurement: object recognition and pose estimation.In recent years,methods based on deep learning have been continuously developed,gradually replacing the method based on manually feature extraction.Point cloud is one of the important representations of 3D data.Compared with 2D images,point cloud contains more information such as depth,and it is representationally simple to be used as input for deep learning models.This thesis is based on the stereo vision technology to obtain the target 3D point cloud,and researches on the object recognition and pose estimation technology based on the 3D point cloud.The main work is as follows:Firstly,depth cameras are compared and binocular camera’s imaging principle is introduced.Then the binocular stereo vision platform is built and calibrated,which collects the point cloud data of the target scene.Because the collected point cloud data has noise and interference,it is necessary to preprocess the point cloud data.In this thesis,conditional filtering,voxel down-sampling and statistical filtering are used to filter the original point cloud to obtain preprocessed point cloud data.Secondly,based on Point Net’s point cloud classification network,this thesis proposes the MV-Point Net network model.The network uses multi-view point cloud data as network input,and uses View-Pooling to aggregate multi-view features.An evaluation experiment was conducted on the model using different View-Pooling methods and numbers of views,and compared with Point Net.The method gains an accuracy rate of 86.5% in the dataset made based on Model Net.The recognition of the real object point cloud experiment was conducted based on the point clouds gained from stereo camera to test the models.Finally,to estimate the pose of the identified object,it is necessary to register the target point cloud with the template point cloud to obtain the transformation matrix.The combination of deep learning coarse registration and ICP fine registration was proposed,whose accuracy is improved compared with the independent algorithms and it is two times faster than the traditional ICP algorithm.更多还原