【作者】 赵明波；

【导师】 付强；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2013， 博士

【摘要】 地面目标检测是空地探测、精确打击领域及自动目标识别技术中的重要研究课题。地面目标容易受各种人工或自然物的遮蔽，这时传统的二维成像探测系统难以获取有效的目标信息。激光成像雷达的波束较窄、方向性好，在一定程度上可以穿透遮蔽物获取部分目标信息，提供反映目标场景几何结构信息的距离像和三维点云数据。虽然激光成像雷达在信息获取上为遮蔽目标探测提供了便利，但遮蔽环境下的成像数据较为复杂，噪声、数据缺失、地物杂波等因素的干扰给遮蔽目标检测提出了挑战。本文针对上述应用背景，按照“数据获取——数据预处理——目标检测”的总体思路，分别从全波形激光成像雷达建模与仿真、激光雷达数据噪声抑制、数据配准和地面遮蔽目标检测算法四个方面展开研究，旨在解决激光成像雷达遮蔽目标检测中的诸多问题，为相关应用奠定基础。绪论部分阐述研究背景和选题意义，总结分析激光成像雷达技术的发展与应用、目标检测技术、遮蔽目标检测方法等领域的研究与发展现状，明确论文的研究内容与思路、结构安排和主要工作。第二章研究全波形激光成像雷达的建模与仿真，解决数据获取问题，为课题相关算法研究提供数据支持。首先阐述全波形激光成像雷达的基本探测原理，明确全波形回波信号的形成过程。然后针对建模与仿真中存在的问题，对仿真系统中的三个主要部分：激光束在目标场景中投影点的获取方法、激光雷达信号与目标场景的作用过程、接收机噪声与距离选通模式，进行重点研究。针对投影点获取过程中存在的大量光线求交运算和传统算法执行效率慢的问题，提出了一种基于激光束空间边界的光线求交快速算法。紧接着设计仿真系统的总体结构、具体实现步骤和系统仿真流程，完成了全波形激光成像雷达仿真系统的构建，可获取多种条件下的成像数据。最后采用理论分析与仿真实验相结合的方法，从全波形回波波形特征、距离选通模式、数据噪声特性三个方面对仿真系统的可信度进行了分析和检验。第三章研究激光成像雷达距离反常噪声抑制算法，为后续处理保障数据质量。首先在分析激光雷达数据各种噪声分布特性的基础上，指出距离反常的影响最大，并结合距离像和三维点云数据中的不同表现形式，分别开展距离反常抑制算法研究。针对距离像数据，依据正常像素在其邻域内至少存在少量的距离相似像素的特点，提出一种基于邻域像素检测的距离反常抑制算法。针对三维点云数据，根据距离正常点和反常点在空间邻域点数量上存在明显的差异性，提出一种基于空间邻域点数量检测的反常点抑制算法。实验验证了所提两种算法在多种条件下的距离反常抑制能力和目标区域数据保护能力，且兼顾去噪性能与执行效率。第四章研究激光雷达数据配准算法，解决遮蔽条件下的数据缺失问题，为实现遮蔽目标检测奠定基础。首先通过对背景需求和多视角拼接算法研究现状的分析总结，得出两视角配准是基础，迭代最近点(Iterative Closed Point，ICP)算法是关键算法的结论。然后结合遮蔽环境下的数据特点，对ICP算法主要步骤的应用策略进行分析和改进，提出了一种基于改进ICP算法的两视角配准算法。在完成两视角配准基础上，提出一种基于积累已配准相邻帧数据的多视角配准算法，该算法在序列配准策略基础上通过合并积累已配准的相邻帧数据来提高待配准数据之间的重叠区域，进而提高配准精度。实验验证了所提两种算法在不同重叠区域、不同配准初值误差条件下均具有较强的鲁棒性和较高的配准精度。最后通过原理与实验分析相结合的手段，对改进ICP算法进行影响因素分析，结果表明配准数据的视角间隔、成像模式影响较大，载体定位、定姿误差影响较小。第五章研究地面遮蔽目标检测算法，实现感兴趣区域(Regions of Interest, ROI)数据的分割提取。在分析遮蔽场景下各种地物空间分布特性的基础上，重点研究了三种空间滤波方法，逐步剔除地物杂波干扰。(1)基于地面数字高程模型(DigitalElevation Model, DEM)估计与高程分割的滤波方法：通过改进形态开运算，提出一种改进的渐进多尺度数学形态学滤波算法，有效实现DEM估计，剔除地面和较高地物上的杂波数据。(2)基于主分量分析法的三维点局部空间分布特性分类检测算法：提取体素邻域内数据的主分量及其贡献率，根据点分布状态与主分量贡献率的关系，实现杂散点、线状点和面状点的分类检测。(3)基于地面阴影分析的目标区域检测算法：利用感兴趣目标对于激光束的不可穿透性和在地面上形成较大面积阴影区域的固有物理特性，通过检测大面积阴影区域来检测目标区域，剔除较小、零散阴影区域内的杂波数据，其中重点推导了算法主要参数与感兴趣目标几何尺寸的关系。在空间滤波基础上，依次采用基于最近距离阈值的谱系聚类法、基于最小外接矩形位置关系的聚类分析法将各个疑似目标区域聚类为各个点集，并结合目标的几何结构信息进一步剔除虚警，实现了ROI数据的分割提取。最后总结出一种基于空间滤波和聚类分析的地面遮蔽目标检测方法的总体框架，为进一步的目标分类、识别提供了基础和条件。第六章总结论文的研究工作及创新性成果，指出存在的问题和后续研究方向。更多还原

【Abstract】 Ground target detection is an important research topic in the field of air-to-groundreconnaissance, precision striking and automatic target recognition. Ground targets arelikely to be obscured by various artificial or natural objects, and it is difficult fortraditional2-D imaging systems to obtain effective target information. The beam ofimaging laser radar (ladar) is narrow and directional. As a result, it can penetrate theoccluder to detect a fraction of the target surface to a certain extent, and obtain the rangeimage and3-D point cloud data which reflect the geometric structure of the target scene.The imaging ladar facilitates information acquisition for obscured target detection, butthe imaging data is more complex in the obscure environment. Therefore, it becomes atechnical problem how to extract target due to the interference of noise, clutter and lackof data. Based on this application background, this dissertation adopts ‘data acquisition,data preprocessing, target detection’ as the overall line of thought, and studies datamodeling and simulation, noise suppression, data registration, and detection algorithm,with the goal to solve the problems of obscured target detection of imaging ladar, andprovides foundation for related applications.Introduction describes the background and significance of the topic considered.The study status of the development and application, target detection technology, andobscured target detection methods for imaging ladar are summarized and surveyed. Ourwork and organization of the dissertation are introduced at last.Chapter2studies the modeling and simulation methods for the full-waveformimaging ladar, which aims at solving the problem of data acquisition and provides datasupport for the related algorithms. Firstly, the basic principle of the full-waveformimaging ladar is described, and the formation process of full-waveform echo signal isrealized. Secondly, focusing on problems in modeling and simulation, we study andmodel the three major components of a simulation system, which are the method ofobtaining projection points of the laser beam in the target scene, the action processbetween the ladar signal and target scene, and the receiver noise and range-gated mode.There are many ray-intersection operations in the obtaining process of projection points,and traditional algorithms are inefficient. To solve this problem we propose a novelray-intersection fast algorithm based on the space boundary of the laser beam. Thirdly,we introduce the overall structure, specific steps and flow chart of the simulation system,and construct a simulation system of full-waveform imaging ladar. The system canprovide imaging data under a variety of conditions. Finally, combining theoreticalanalyses and simulation experiments, we analyze and verify the reliability of thesimulation system from three aspects, which are waveform characteristics of thefull-waveform echo signal, range-gated mode, and noise characteristics of the imaging data.Chapter3studies the noise suppression algorithm of ladar data to provide highquality data for subsequent processing. Firstly, based on analysis of the distributioncharacteristics of all kinds of noise in the ladar data, we point out that the effect of rangeanomaly is the most important factor. Then combining different behaviors of rangeanomaly in the range image and3-D point cloud data, the suppression algorithms arestudied separately. For the range image, we propose a range anomaly suppressionalgorithm based on neighborhood pixels detection, according to the fact that there are atleast a small amount of range-similar pixels in the neighborhood of the normal pixel.For the3-D point cloud data, there are significant differences in the number of points inspace neighborhood (NPSN) between the normal and anomalous point, and ananomalous point suppression algorithm based on NPSN detection is proposed. Wevalidate the range anomaly suppression performance and the capability of protectingtarget data of the two proposed algorithms by experiments, taking into account theirdenoising performance and implementation efficiency.Chapter4studies the ladar data registration algorithm, which is to solve theproblem of data insufficiency in the obscure environment and provides foundation forobscured target detection. Firstly, through analyzing and summarizing the backgroundrequirements and the study status of multi-view registration algorithm, we find that thetwo-view data registration is fundamental and the iterative closest point (ICP) algorithmlies in the core. Secondly, combining the data characteristic in the obscure environment,we analyze and improve the application strategies of the ICP’s main steps, and propose atwo-view data registration algorithm based on improved ICP. Thirdly, a multi-view dataregistration algorithm based on aggregating the adjacent frames which are alreadyregistered is proposed. Based on the sequence registration strategy, it increases theoverlap region between the pending registration frames by aggregating, and furtherimproves the registration accuracy. Experiment results validate that the two proposedalgorithms have greater robustness and higher registration accuracy under conditions ofdifferent overlap regions and initial errors. Finally, we analyze the impact factors of theimproved ICP through the means of a combination of theoretical and experimentalanalysis. The results show that the effect of view-interval and imaging mode of theregistration data is more significant, and the effect of the carrier position and attitudeerrors is less.Chapter5studies the ground obscured target detection algorithm for segmentingand extracting the data in the region of interest (ROI). Based on the analysis of spatialdistribution characteristics of various ground-objects in the obscure environment, wefocus on three spatial filtering methods to filter out the clutter data step by step. Firstly,we study a filtering method based on the estimation of ground digital elevation model(DEM) and segmentation by elevation threshold to filter out the clutter data on the ground and higher objects. To reliably estimate the DEM, a filter based on improvedprogressive multi-scale mathematic morphology is proposed by improvingmorphological opening operation. Secondly, we propose a classification-detectionalgorithm of the local spatial distribution characteristics of3-D point based on theprincipal component analysis (PCA). It extracts the principal component andcontribution rate of the data in voxel neighborhood, then classifies and detects thescattered points, linear points and planar points according to the relationship of pointdistribution state and contribution rate of principal component. Thirdly, we propose amethod of target region detection based on ground shadow analysis to filter out theclutter data of interference objects which cannot form ground shadow of a large area.According to the inherent physical characteristics of the targets of interest that they arenot penetrable to laser beam and can form ground shadow of a large area, the targetregion is detected by detecting the ground shadow of a large area. We derive therelationship of the algorithm’s main parameters and the geometrical dimensions of thetarget of interest. Based on the spatial filtering results, the sets of data points for eachsuspected target region are obtained by orderly using the hierarchical clustering methodbased on closest distance threshold and clustering analysis method based on positionalrelationship of minimum bounding rectangle. By combining with the priori geometryinformation of target, false alarms are further removed, and the3-D point cloud data ofROI is obtained. Finally, we sum up a general framework of the detection method forground obscured target, to provide a foundation for the subsequent target classificationand identification.Chapter6summarizes the research work and innovative achievements of thedissertation. The shortcomings and future work are also included.更多还原