Generation of stereoscopic image sequences using structure and rigid motion estimation by extended Kalman filters

Abstract

In this paper an object-based method to generate additional stereo views from monoscopic image sequences using rigid motion and structure estimation by extended Kalman filters is presented. First, rigid scene objects are segmented and feature points in each object are extracted and tracked throughout the video sequence. Then, motion, structure and focal length are estimated recursively for each object, using extended Kalman filters, as described in Azarbayejani and Pentland (1995). Furthermore, the feature points and depths in the stereo image are computed and interpolated using 2-D Delaunay triangulation. Finally, a stereo image generation algorithm is proposed that uses the camera and structure equations to project the 3-D points in a new virtual stereo view for each frame. The generation of stereoscopic scenes is possible even when multiple moving rigid objects exist in the scene. Experimental results show that a layered stereo object representation yields improved results.