USING THE ORTHOGONAL PROJECTION FOR PARAMETER INITIALIZATION IN THE 3D RECONSTRUCTION OF DISTANT OBJECTS
Keith F. Blonquist, Robert T. Pack · 2010
We present a method for the 3D reconstruction of objects from images captured in a surveillance setting using uncalibrated hand-held cameras. The images considered are of distant objects occupying a small portion of the field-ofview (FOV); have been captured in a non-systematic manner from varying aspects, distances, and angles; and have no a priori target information for 3D coordinates or scale. The current algorithms used to estimate camera orientations and perform 3D reconstruction rely to some degree on initial approximations of camera orientation and/or 3D coordinates of the target object or control points. Parameter initialization in most surveillance applications is problematic due to the non-regular camera arrangements and lack of target information. Using some of the strengths of the orthogonal model, this paper presents a robust initial approximation method suited to the demands of surveillance imagery. The method assumes initially that the images are orthogonal, allowing for a direct linear solution without the necessity of initial approximations. The accuracy of the approximations from the orthogonal method can then be improved using a “perspective-corrected” orthogonal projection that incorporates the approximate data in order to adjust the images from perspective images to nearly orthogonal images. Finally, the “perspective-corrected” orthogonal solution is used to approximate the camera orientations, and a standard bundle adjustment based on the perspective projection can be carried out. The reliability of the method using varying imaging distances and cameras is demonstrated through experimentation with images taken in the field.