【作者】 王茁；

【导师】 陆军；

【作者基本信息】 哈尔滨工程大学 ， 控制科学与工程， 2021， 硕士

【摘要】 计算机视觉技术是一门研究机器如何“看”到世界的技术。具体来讲,是通过诸如相机(Camera)、毫米波雷达(Radar)和激光雷达(Lidar)等传感器将目标的视觉信息采集到计算机,并对这些数据进行一系列的处理,并实现识别、检测、分割等功能。识别技术指的是机器通过传感器“发现”目标,并通过一些列计算、分析得出目标的种类、方向、速度等信息。识别技术广泛应用于各个领域,如制造业领域、安防领域、机器人领域、交通领域、军事领域等。识别技术根据传感器采集到的数据的不同可以划分为图像识别、点云识别、多传感器融合识别等。近年来,随着深度学习技术迅猛发展,目标识别、检测、分割等领域也得到了快速的发展。本文主要工作分为三大部分:首先,根据点云的分辨率自适应地调整滤波器参数,提出一种基于预关键点邻域曲率均值的关键点查找算法。该算法不依赖于单点曲率值,增强了对噪声的鲁棒性,降低相同局部区域上的关键点重复性。提出一种与其邻域各点之间基于距离和重心的法向关系的特征描述符,改善了面对噪声点的鲁棒性和保证了描述符的唯一性。通过配准实验验证了特征描述符的有效性。所提出的特征描述符在配准实验中具有良好的准确度,计算效率和抗噪声能力。然后,介绍为了构建复合体素特征的原始点云处理方法,通过删除部分点并进行循环体素滤波来为后续的步骤降低计算量。介绍对于处理后的点云进行体素构建的方法,通过按照一点的比例对于处理后的点云数据进行体素划分。详细描述体素构建过的程,对每个体素中的所有点进行采样、深度特征提取,对体素中的所有点进行关键点查找,对于关键点进行局部曲率直方图的特征描述符构建,将局部曲率直方图特征融入深度体素特征,对于复合特征进行池化降维。对该复合特征进行理论上的解释和可行性的分析。最后,介绍完整网络的识别流程,在特征图构建后,通过修改后的RPN网络对目标进行检测和识别。介绍RPN网络的结构和使用情况,通过多个3D卷积核对特征进行采样,最终获得位置回归图和概率预测图。介绍损失函数的设计。介绍KITTI数据集在训练中的使用情况以及训练网络的细节。通过对比实验,验证该网络的性能。更多还原

【Abstract】 Computer vision technology is a technology that studies how machines "see" the world.Specifically,the visual information of the target is collected to the computer through sensors such as camera,millimeter-wave radar,and lidar,and a series of processing of these data is performed to realize recognition,detection,segmentation and other functions.Recognition technology means that the machine "discovers" the target through sensors,and obtains the target’s type,direction,speed and other information through a series of calculations and analysis.Recognition technology is widely used in various fields,such as manufacturing,security,robotics,transportation,military,etc.Recognition technology can be divided into image recognition,point cloud recognition,multi-sensor fusion recognition,etc.according to the data collected by sensors.In recent years,with the rapid development of deep learning technology,the fields of target recognition,detection,and segmentation have also developed rapidly.The main work of this paper is divided into three parts:First,the filter parameters are adjusted adaptively according to the resolution of the point cloud,and a key point search algorithm based on the mean value of the curvature of the pre-key point neighborhood is proposed.The algorithm does not depend on the curvature value of a single point,which enhances the robustness to noise and reduces the repetitiveness of key points on the same local area.A feature descriptor based on the normal relationship between the distance and the center of gravity between each point in its neighborhood is proposed,which improves the robustness to noise points and ensures the uniqueness of the descriptor.The registration experiment verifies the effectiveness of the feature descriptor.The proposed feature descriptor has good accuracy,computational efficiency and anti-noise ability in registration experiments.Then,the original point cloud processing method for constructing composite voxel features is introduced,which reduces the amount of calculation for subsequent steps by deleting some points and performing loop voxel filtering.The method of constructing voxels for the processed point cloud is introduced,and the processed point cloud data is divided into voxels according to the ratio of one point.The process of voxel construction is described in detail,all points in each voxel are sampled,deep feature extraction is performed,key point search is performed on all points in the voxel,and feature descriptors of the local curvature histogram are performed for key points Construct,integrate local curvature histogram features into deep voxel features,and perform pooling and dimensionality reduction for composite features.The composite feature is theoretically explained and feasibility analysis.Finally,the identification process of the complete network is introduced.After the feature map is constructed,the target is detected and identified through the modified RPN network.Introduce the structure and usage of the RPN network,sample features through multiple 3D convolution kernels,and finally obtain a position regression map and a probability prediction map.The design of the loss function is introduced.Introduced the use of KITTI data set in training and the details of training network.Through comparative experiments,the performance of the network is verified.更多还原