Abstract
This paper describes a method for analyzing optical flow and reconstructing 3-D motion and structure using three views. Instead of taking correspondence between tokens in stereo images, it tries to estimate motion and structure by comparing the analytic representation of optical flows in stereo images. The assumptions made here are: 1) images are formed by orthographic projection, 2) a rigid motion is to be analyzed. Firstly it is shown that the optical flow of an arbitrary motion is that of purely rotational motion under orthographic projection. Then by combining the parameters describing apparent rotational optical flow observed from each view, the 3-D velocity vector field can be reconstructed. The calculation of depth map of the object in motion follows directly. In order to adjust the estimated parameters of the rotational optical flows from three views and obtain a more plausible solution, an optimizing process is proposed where the geometrical relation between three views are used as constraints. The results of applying this method to computer generated optical flow and real optical flow are shown and the effectiveness of the method is tested. The merit of the method is that it can recover the 3-D velocity vector field and the depth map of the moving object from noisy optical flow avoiding correspondence problem peculiar to stereoscopic techniques.
