Detecting Correspondences between Video Image Frames and Upgrading Scene Analysis Using Them

1.图像间的对应性检测我们提出了一种新的方法，用于自动提取两幅图像中的特征点，并自动检测它们之间的对应性，结合基于几何约束和全局一致性条件的投票方案。我们通过设计一种技术来补充它们，该技术用于使用模板匹配来生成更密集的特征对应，并且该技术用于使用尝试性的3-D重建来检测不匹配。2.视频图像中的移动对象的分离我们创建了一种新的方法，用于通过估计缺失部分来扩展部分跟踪通过视频流的特征点轨迹。我们还创建了一个有效的方法，用于分离跟踪轨迹的背景部分和独立移动的对象部分。我们还提出了一种技术，用于检测移动对象区域，通过估计运动的背景和减去它从个别视频帧。3.3-D重建从imagesWe设计了强大的技术，从视频流和两个单独的图像的3-D重建。我们还得出了一个强大的技术，从一个单一的图像，利用知识的平行性和正交关系的场景中的3-D重建。4.几何推理的最优估计和模型选择在最优参数估计和模型选择的数学分析方面取得了相当大的进展，并讨论了模型选择准则的含义，称为几何AIC和几何MDL，以及理论精度界，所谓的KCR下限，主要研究者提出的是明确的。

1.Correspondence detection between imagesWe proposed new methods for automatically extracting feature points in two images and automatically detecting correspondences between them, combining voting schemes based on geometric constraints and global consistency conditions. We supplemented them by devising a technique for generating denser feature correspondences using template matching and a technique for detecting mismatches using tentative 3-D reconstruction.2.Separation of moving objects in video imagesWe created a new method for extending the feature point trajectories partially tracked through a video stream by estimating the missing parts. We also created an effective method for separating the tracked trajectories into a background part and an independently moving object part. We also proposed a technique for detecting moving object regions by estimating the motion of the background and subtracting it from the individual video frames.3.3-D reconstruction from imagesWe devised robust techniques for 3-D reconstruction from a video stream and from two separate images. We also derived a robust technique for 3-D reconstruction from a single image using knowledge about parallelism and orthogonality relations in the scene. Furthermore, we created a new technique for displaying the reconstructed 3-D shape in such a way that its shape characteristics are preserved.4.Optimal estimation and model selection for geometric inferencesA considerable progress is made in mathematical analysis of optimal parameter estimation and model selection, and the meaning of the model selection criteria, called the geometric AIC and the geometric MDL, and the theoretical accuracy bound, called the KCR lower bound, that the principal investigator proposed is made clear.