• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文学位 >Pattern modeling and classification in vision systems.
【24h】

Pattern modeling and classification in vision systems.

机译：视觉系统中的模式建模和分类。

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

This thesis presents new methods and algorithms for a number of related computer vision tasks: foreground-background segmentation with fixed or moving camera, object tracking, sequential face detection, and illumination correction. These topics are of particular importance for visual surveillance systems. These tasks are mostly approached by modeling and classifying patterns. So the performance depends largely on how accurately the patterns are modeled and how effectively the patterns can be classified by the algorithms.; The first part of the thesis introduces a novel transform domain approach for foreground segmentation with a static camera. The use of the DCT features and a simple temporal prediction scheme results in robust segmentation with fast computation. It is also able to handle some typical difficulties in these problems, including illumination change and repetitive background motion. In the second part of the thesis, a histogram-based appearance model is employed for tracking based on the segmentation result. To make the tracking more robust to camouflage and partial occlusion, an irregularly-shaped feature selection and confidence propagation scheme is introduced. The third part deals with segmentation with a moving camera, where the knowledge of the motion flow and the focus of expansion are exploited to estimate the 3D camera ego-motion. With the estimated 3D motion, the segmentation is carried out by detecting the motion flow outliers. Different noise models are analyzed and the corresponding maximum likelihood estimates are compared. The forth part of the thesis studies a fast face detection system using sequential detection to trade off detection accuracy with computation time. The detection consists of a feature point detection step, a subregion-based Bayesian detector, and a wavelet-based Bayesian detector. This part also includes a discussion of the general discriminant analysis methods, discriminative feature selection, and Bayesian error reduction. The last part of the thesis presents a general illumination correction algorithm that is designed to enhance the local contrast on images with a large dynamic range. An affine model is employed to simulate the multiplicative and additive effects of illumination, and a fast low pass filtering method is proposed to estimate the affine model parameters.
机译：本文提出了用于许多相关计算机视觉任务的新方法和算法：使用固定或移动相机进行前景-背景分割，对象跟踪，顺序人脸检测以及照明校正。这些主题对于视觉监控系统特别重要。这些任务主要通过对模型进行建模和分类来解决。因此，性能很大程度上取决于对模式进行建模的准确性以及算法对模式进行分类的有效程度。论文的第一部分介绍了一种用于静态摄像机前景分割的新型变换域方法。 DCT功能和简单的时间预测方案的使用可实现快速计算的鲁棒分割。它还能够处理这些问题中的一些典型困难，包括照明变化和重复的背景运动。在论文的第二部分，基于直方图的外观模型被用于基于分割结果的跟踪。为了使跟踪对迷彩和部分遮挡更加鲁棒，引入了不规则形状的特征选择和置信度传播方案。第三部分处理运动相机的分割，其中利用运动流的知识和扩展焦点来估计3D相机的自我运动。利用估计的3D运动，通过检测运动流离群值进行分割。分析了不同的噪声模型，并比较了相应的最大似然估计。论文的第四部分研究了一种快速的人脸检测系统，该系统使用顺序检测来权衡检测精度与计算时间。该检测包括特征点检测步骤，基于子区域的贝叶斯检测器和基于小波的贝叶斯检测器。本部分还讨论了一般判别分析方法，判别特征选择和贝叶斯误差减少。本文的最后一部分提出了一种通用的照度校正算法，该算法旨在增强动态范围较大的图像的局部对比度。仿射模型被用来模拟照明的乘法和加法效应，并提出了一种快速的低通滤波方法来估计仿射模型参数。

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