【作者】 陈晓清；

【导师】 庄钊文；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2010， 博士

【摘要】 随着激光技术的发展,激光成像雷达在现代战争复杂战场环境中逐渐获得了广泛的应用,其相关目标识别技术已成为国内外研究的热点问题。本文基于激光成像雷达距离像,围绕噪声抑制、图像分割和目标分类与识别这三个方面展开深入研究。论文的主要内容如下:绪论部分首先阐述了论文的研究背景及意义,然后概述了激光成像雷达目标识别的研究现状及存在的问题,最后介绍论文的主要工作。第二章研究了距离像噪声抑制。针对距离像中距离反常的形成机理和噪声特性,提出了三种距离反常噪声抑制算法。(1)提出了基于排序差分和自适应中值滤波的距离反常噪声抑制算法,通过对大滤波窗口内的像素值进行排序差分,选择低于门限的最大连续差分值所对应的像素值作为距离正常值,实现对距离反常噪声的自适应中值滤波。(2)提出了基于最优参数自适应中值滤波的距离反常噪声抑制算法,通过计算当前像素值与中值的差,再与两个门限比较来获取滤波结果。(3)提出了基于包围准则的距离反常噪声抑制算法,根据包围准则检测噪声,结合中值滤波和加权均值滤波,达到噪声抑制目的。实验结果表明,这三种算法都能够有效抑制距离反常噪声,同时保护目标的边缘细节信息。第三章研究了距离像图像分割。针对经典边缘检测算子在检测距离像边缘与分割距离像时的功能局限,提出了两种改进算法。(1)提出了基于Canny算子的深度区域边缘检测算法,先利用排序差分初步分割,再采用Canny算子检测边缘信息并修正,提取出完整的目标轮廓边缘。(2)提出了基于边缘控制区域生长的图像分割算法,自主选取目标区域的生长起始点,并利用目标轮廓边缘控制区域生长过程,实现了完整目标区域的图像分割。实验结果表明,将两种算法组合应用于激光雷达距离像,能够收到较好的图像分割效果。第四章研究距离像目标分类与识别。基于图像特征和模型匹配两种技术途径提出了几种距离像识别算法。基于图像特征的技术途径,提出了四种距离像识别算法:(1)提出了基于奇异值特征的距离像识别算法,通过分析识别率与奇异值特征数目的关系,找出有效特征并应用最优参数支持向量机进行分类识别。(2)提出了基于小波变换的距离像识别算法,先对距离像进行二维小波变换,然后从近似分量和细节分量中提取分类特征,应用最优参数支持向量机进行分类识别。(3)提出了基于矩特征的距离像识别算法,分别从距离像中提取Hu矩和Zernike矩特征,作为输入矢量训练RBF神经网络实现分类识别。(4)提出了基于纹理特征的距离像识别算法,通过计算距离像的灰度共生矩阵,提取出13个纹理特征作为输入矢量训练RBF神经网络用于分类识别。根据模型匹配的原理,提出了一种基于模型匹配的距离像识别算法,包括矩形估计、模型匹配和相似度测量三个主要步骤:(1)提出了SLEC矩形估计算法,通过将点云栅格化得到二值图像并以形态学闭运算填充内部孔洞,采用Sobel算子提取边界并用Hough变换得到其角度参数,以边界直线的角度为约束条件寻找面积最小矩形得到目标的方向和尺寸估计。(2)提出了一种改进迭代最近点算法,将标准ICP算法点与点之间的匹配扩展到点与面之间的匹配,通过寻找达到点云和CAD模型之间欧氏距离之和最小的刚体变换实现精确匹配。(3)提出了一种点云和CAD模型之间的目标相似度测量方法,将归一化欧氏距离作为相似性的测度,据此实现目标的分类识别。仿真实验表明,这几种距离像识别算法均取得了较好的识别效果。结束语总结了论文的创新性成果,并给出下一步的研究方向。更多还原

【Abstract】 With the development of laser technology, laser imaging radar was gradually gained wide application in complicated battlefield of modern warfare. The target recognition technology for laser imaging radar has become a hot issue at home and abroad. This dissertation researches on the target recognition technology for laser imaging radar range image, which includes the noise reduction, the image segmentation, and the target classification and recognition. The main contributions of the dissertation are demonstrated as follows:In the introduction, firstly, the research background and significance of this dissertation are set forth. Then the research state and problem of laser imaging radar target recognition are summarized. Finally, the main work of this paper is introduced.In chapter 2, the noise reduction of range image is studied. Based on the formation mechanism and noise characteristic of rang anomalies in range image, three types of noise reduction algorithm is proposed. (1) A range anomalous suppression algorithm based on differential sorting and adaptive median filter is proposed. Firstly, the sorting and difference are carried out for pixels with big filtering window. Then, the pixels corresponding to the maximum continuous difference value below threshold are regarded as normal range values. Finally, a self-adaptive median filtering is carried out for range anomalies. (2) A range anomalies reduction algorithm based on parameter optimization adaptive median filter is proposed. The filtering value is achieved by calculating the difference between present pixel and median and comparing the difference with two thresholds. (3) A range anomalies reduction algorithm of laser imaging radar range image based on surrounding criterion is proposed. Firstly, the noise is detected based on surrounding criterion. Then using a combining median filter with weighted mean filter, the noise reduction is carried out on range image. The experimental results show that the range anomalies noises can be efficiently suppressed based on these three algorithms and the detail of object in range image is well protected.In chapter 3, the image segmentation of range image is studied. According to the functional limitations of classical edge detection operators in detecting the edge of range image and segmenting the range image, two improved algorithm are proposed. (1) An edge detection method of depth region based on Canny operator is proposed. Firstly, the initial segmentation is carried out by using sort differential operator. Then, the edge detection and correction is carried out based on Canny operator. Finally, the contour edge of object is extracted completely. (2) An image segmentation method based on edge control region growth is proposed. Firstly, the initial pixel of edge region growth is automatically selected, and then the region growth process is dominated based on object contour edge. Finally, the image segmentation of integral target region is realized. Experimental results show that combining these two methods can gain good image segment effect of laser imaging radar range image.In chapter 4, the target classification and recognition of range image is studied. According to two technical approaches (image features and model matching), some recognition algorithms of range image are proposed. Based on image features, four recognition algorithms of range image are proposed. (1) A recognition algorithm of range image based on singular value feature is proposed. According to the relationship between recognition rate and the number of singular value feature, important features are obtained. Then, target recognition is carried out using the support vector machines of optimal parameter. (2) A recognition algorithm of range image based on wavelet transform is proposed. Firstly, two dimensional wavelet transform of range image is carried out. Then features are extracted from approximate part and detail part. Finally the target is identified using the support vector machines of optimal parameter. (3) A recognition algorithm based on moment feature is proposed. Firstly, Hu moment and Zernike moment features of range images are extracted separately. Then the target is identified using the RBF neural network. (4) A recognition algorithm based on texture feature is proposed. Thirteen texture features based on Gray Level Co-occurrence Matrix are extracted from range image. Then the target is identified using the RBF neural network. According to model matching principle, a recognition algorithms of range image is proposed, including three major steps: rectangle estimation, model matching and similarity measure. (1) A SLEC rectangle estimation algorithm is proposed. Firstly, a binary image is got by rasterizing point cloud and the internal holes are filled through morphological closing operation. Then, the edge is extracted based on Sobel operator and angle parameters are got by Hough transform. Finally, the target direction and size estimation are got by searching the smallest rectangle with constraint of the boundary line angle. (2) A modified ICP algorithm is proposed. The standard ICP algorithm only matches point to point, while the modified ICP algorithm is expanded to matching point and surface. The accurate matching is carried out a rigid body transformation which searches the minimum euclidean distance between the point cloud and CAD models. (3) To measure the similarity between target point cloud and CAD models, a measurement algorithm is proposed which using the normalized euclidean distance as similarity measure. The target recognition is based on this similarity measure. Experimental results show that these recognition algorithms have good recognition results.Finally, the innovation of this paper is summarized. The last chapter also discusses the future work to be further researched.更多还原