Decomposition of transformation matrices for robot vision

The relationship between the three-dimensional coordinates of a point and the corresponding two-dimensional coordinates of its image, as seen by a camera, can be expressed in terms of a 3 by 4 matrix using the homogeneous coordinate system. This matrix is known more generally as the transformation matrix and it is well known that such a matrix can be determined experimentally by measuring the image coordinates of six or more points in space, whose three-dimensional coordinates are known.Such a transformation matrix can be derived analytically from knowledge of the camera position, orientation, focal length and scaling and translation parameters in the image plane. However, the inverse problem of computing the camera location and orientation from the transformation matrix involves solution of simultaneous nonlinear equations in several variables and is considered difficult.In this paper we present a new and simple analytical technique that accomplishes this inversion rather easily. This technique is quite powerful and has applications to a wide variety of problems in Computer Vision for both static and dynamic scenes. The technique has been implemented as a C program running under Unix and works well on real data.