Abstract
Typical methods for camera calibration and image rectifica- tion from a single view assume the existence of straight parallel lines from which vanishing points can be computed, or orthogonal structure known to exist in the scene. However, there are practical situations where these assumptions do not apply. Moreover, from a single family of paral- lel lines on the ground plane there is insufficient information to recover a complete rectification. Here we study a generalization of these meth- ods to scenes known to contain parallel curves. Our method is based on establishing an association between pairs of corresponding points lying on the image pro jection of these curves. We show how this method can be used to compute a least-squares estimate of the focal length and the camera pose from the tangent lines of the associated points, allowing complete rectification of the image. We evaluate the method on high- way and sports track imagery, and demonstrate its accuracy relative to a state-of-the-art vanishing point method.