【作者】 李世飞；

【导师】 沈振康；

【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】 在最新的激光成像精确制导技术中,利用三维激光雷达图像获取的三维目标数据进行三维目标识别是需要解决的一个关键问题。在三维视觉理论中,三维目标识别是一个极富挑战性的难题,本文紧扣三维目标识别这个核心问题,对基于深度图像的三维目标识别技术进行了细致深入的研究。本文的主要工作和创新成果如下:(1)深度图像的表示是实现三维目标识别的前提。在研究当前各种深度图像获取手段的基础上,提出了将深度图像分为伪灰度表示、网格表示和点云表示的分类方法,并对各种深度图像表示方法进行了分析。(2)根据伪灰度表示深度图像的特性,研究了深度图像的曲面特性的直接计算、数值估计和曲面拟合等几种方法。针对现有拟合方法均为局部拟合的特点,创新地提出了利用移动最小二乘方法实现深度图像整体拟合的方法,得到待拟合曲面的一种隐式整体表达,仿真试验结果表明这种整体拟合方法不仅具备很好的拟合精度,而且便于对曲面特性进行计算。在此基础上,推广得到了基于最小二乘的曲面拟合一般性模型框架。该模型框架不仅对已有的各种不同拟合方法进行了本质上的理论区分,也为探索新的深度图像拟合方法指明了道路。(3)针对现有三维目标表达方法无法有效解决自由形态物体整体表达的问题,提出了利用形状索引直方图和距离直方图这样的整体特征来进行三维目标识别的新方法,构建了基于形状索引直方图的目标识别算法流程。仿真结果表明,该方法可以有效避免了计算代价极大的特征点对的对应性搜索,并且能够较好地满足三维目标识别中的旋转和尺度不变性。(4)为了解决三维目标姿态多样性所导致的不变量难以实现的问题,提出了三维目标识别中的相关不变性这一全新概念。并根据通过整体识别整体的思想提出了基于Tsallis熵和旋转图像的三维目标识别方法,首先利用旋转图像将三维目标识别问题映射成二维问题,然后采用Tsallis熵对旋转图像特征进行表达,实现了数据的急剧衰减,避免了繁琐的特征点对匹配搜索,并在理论上证明的该方法的平移、旋转和缩放变换下相关不变性,理论分析和实验数据表明该方法具有极高的三维目标识别正确率。更多还原

【Abstract】 Making use of three-dimensional target data obtained from the three-dimensional laser radar to recognize three-dimensional object is the core process of lasted precision guided weapon technology. In three-dimensional computer vision, three-dimensional object recognition plays a prominent role and is a hard and challenging task. This thesis takes three-dimensional object recognition as the study’s core and has a detailed in-depth study to three-dimensional object recognition based on range images. The main work is summarized as follows:(1) In three-dimensional computer vision range image’s representation is the base of three dimensional object recognition. These thesis summaries the range image’s representation as three types, these are pseudo-gray representation, gird representation and point cloud representation.(2) According to the properties of range image with pseudo-gray representation, this thesis studies three methods of calculation of curvature properties, they are respectively direct calculation method; Numerical estimates method and Surface Fitting methods. since the existing fitting methods are local fitting methods, this thesis presents a new global surface fitting scheme by using moving least squares methods. Compared with the standard least squares, moving least squares method can obtain an implicit general representation and better fitting accuracy. This thesis generalizes the moving least squares methods and deduces a general model frame of surface fitting. This frame not only demonstrates the Theoretical difference of all the surface fitting methods and also indicates a direction to design new surface fitting method.(3) As we all know, the existing representations of free-form 3D objects are local representations .In view of this, this thesis proposes a new three-dimensional object recognition method based on shape index histogram and range histogram. This method not only avoids the search of characteristic corresponding points and satisfies to a certain extent rotation invariance, translation invariance and scale invariance.(4) In three-dimensional computer vision constructing pure invariant is very difficult. In view of this, the paper defines a correlation invariability with a relaxative standard and proposes a new three-dimensional object recognition method based on Tsallis entropy and spin images. This method achieve the rapid decay of data and avoid search of characteristic corresponding points, moreover this representation possesses rotation invariance, translation invariance and scale invariance. Theoretical analysis and simulations prove this method’s high performance in three dimensional object recognition.更多还原