Estimation of 3D structure and camera calibration from images under noisy conditions

En savoir plus sur le livre

Extracting structural information from images or videos is crucial for technologies like 3-D television, driver assistance systems, and surveillance. For Euclidean 3-D reconstruction, a dense motion field and camera parameters must be identified. This work aims to automatically estimate point correspondences and camera parameters from a stereo image pair simultaneously. It combines camera calibration and stereo matching in an iterative process. Epipolar geometry plays a key role by encapsulating camera parameters and reducing the search effort in point correspondence analysis. Since epipolar geometry is determined from point correspondences, addressing noise and outliers is essential. A distance measure based on the uncertainty of the epipolar geometry is introduced for this purpose. Additionally, a polynomial method is derived to estimate focal length by exploiting constraints on the essential matrix. To achieve dense correspondence analysis, homographies are derived from point correspondences within color-segmented regions. Coupled with Delaunay triangulation, these homographies serve as predictors for further correspondence searches, resulting in dense motion estimation. Image pairs captured over time may include moving objects, which do not adhere to the epipolar constraint and cannot aid in camera calibration. An enhanced framework is presented to separate static and dynamic scene components effectively.