• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文学位 >Parameterized surface models for binocular stereo vision.
【24h】

Parameterized surface models for binocular stereo vision.

机译：用于双目立体视觉的参数化表面模型。

获取原文
获取原文并翻译 | 示例
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

Described in this thesis is a framework for reconstructing three-dimensional surfaces from stereo image intensity data. Given a pair of stereo intensity images, the task is to segment the scene into regions, each region being defined by a rudimentary set of parameters. We take the novel approach to data fusion known as "the method of competitive priors." Data fusion methods typically have a single prior model with different sensory channels contributing to the model. For example one may combine stereo with shading to recover 3D information. Here we have a single channel of information but allow a number of prior models to compete to fit the data. A prior probability distribution is assigned for each type of surface and points that are not "on the surface" are modeled as coming from a uniform distribution. We obtain an energy functional from formulating the probabilities with Gibbs' distributions. The maximum a posteriori (MAP) estimate which best models the scene is selected by simultaneously determining the most likely prior and the model parameters. The points that lie on the surface are then recovered.; This model is then extended to one which allows for multiple surfaces in an image pair and which also demands surface coherence. The first step in this expanded algorithm is to roughly calculate disparity values at each point based on a windowing scheme and a nonlinear filter. This preprocessed image is then input into a system combining region competition (54) with stereo. A number of regions are distributed throughout the scene, and these regions grow based on how well they fit parameterized surfaces. After this process is iterated the regions are then allowed to merge based on an overall energy functional. The method segments the scene at each step, with a parameterized surface defined in each region. The resulting description is compact and is useful as a starting point for an active vision system which integrates information from multiple viewpoints; as a result, a robust and stable description of the world, less prone to viewpoint dependency then those based on the current viewpoint is produced.
机译：本文描述的是一种从立体图像强度数据重建三维表面的框架。给定一对立体声强度图像，任务是将场景划分为多个区域，每个区域由一组基本参数定义。我们采用称为“竞争先验方法”的新颖数据融合方法。数据融合方法通常具有单个先验模型，而不同的感官通道对该模型有所贡献。例如，可以将立体声与阴影相结合以恢复3D信息。在这里，我们只有一个信息渠道，但允许许多现有模型竞争以拟合数据。为每种类型的表面分配一个先验概率分布，将不在“表面上”的点建模为来自均匀分布。我们通过用吉布斯分布来表示概率来获得能量函数。通过同时确定最可能的先验和模型参数，最大后验（MAP）估计值可以选择对场景进行最佳建模的模型。然后恢复位于表面上的点。然后将此模型扩展到允许在一个图像对中使用多个表面并且还需要表面一致性的模型。这种扩展算法的第一步是基于加窗方案和非线性滤波器来大致计算每个点的视差值。然后将该经过预处理的图像输入到将区域竞赛（54）与立体声相结合的系统中。许多区域分布在整个场景中，这些区域基于它们对参数化曲面的拟合程度而增长。重复此过程后，然后允许区域基于整体能量函数进行合并。该方法在每个步骤中使用在每个区域中定义的参数化曲面对场景进行分割。所得到的描述是紧凑的，并且可以用作主动视觉系统的起点，该系统集成了来自多个视点的信息。结果，产生了对世界的鲁棒和稳定的描述，与基于当前观点的观点相比，不那么容易出现观点依赖。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号