Camera calibration from silhouettes

Abstract

Silhouettes are often dominant image features, and they provide valuable information about both camera properties and object motion. However, since silhouettes are viewpoint dependent, few direct correspondences can be obtained between them. This makes camera calibration from silhouettes a challenging problem. This thesis develops new algorithms for camera calibration using silhouettes in an image sequence, with the objective of using the camera calibration to compute the motion and hence achieve a Euclidean reconstruction. The thesis first studies the problems of camera calibration using silhouettes of a sphere. It is shown that the common pole and polar with respect to the conic images of two spheres are also the pole and polar with respect to the image of the absolute conic (IAC). This pole-polar relationship provides two linear constraints for estimating the IAC and hence allows a camera to be linearly calibrated from an image of three spheres. The thesis then studies the problem of self-calibration for a turntable sequence. By exploiting the parametrization of the fundamental matrix, a fixed scalar which can be obtained from the epipoles in an image triplet is found to be a crucial element in both calibration and motion estimation. The imaged circular points can then be formulated in terms of all image invariants and the scalar and used to recover the camera intrinsics. The camera motion and a Euclidean reconstruction then follow. However, this method relies on an input of a dense image sequence for initializing the image invariants. A possible solution for a sparse sequence is to introduce more image features (points) into the system, and a novel method using silhouettes and an image point track is developed accordingly.