Epipolar geometry in stereo, motion and object recognition. A unified approach (Q1379185)

Two linear mappings \(p\mapsto p'\) and \(p\mapsto p''\) of projective 3-space P\(^3\) to the projective plane P\(^2\) have the following property: In general there are points \(z'\), \(z''\) and a one-to-one linear mapping \(\phi\) of the line pencil \(z'\) onto the line pencil \(z''\) such that for a \(q\in\text{P}^3\) the line \(z'q'\) is mapped by \(\phi\) to \(z''q''\). This is well known in descriptive geometry and photogrammetry [see for example \textit{E. Müller}, Vorlesungen über darstellende Geometrie, Vol. 1 (Leipzig, 1923), or \textit{H. Brauner}, Abh. Math. Semin. Univ. Hamb. 53, 154-169 (1983; Zbl 0519.51003)]. In Computer Vision, this fact is called the \textit{epipolar geometry} of the two images. The authors, deliberately avoiding the language of projective geometry in order to make the book more accessible to the Computer Vision community, study a list of problems from the viewpoint of epipolar geometry. These include different camera models and their epipolar geometry; the recovery of the epipolar geometry from the images, using either point or line information; the problems of stereo, motion and object recognition.