Cylindrical object reconstruction with a moving camera embedded in a "poor" robotic platform

Evolution of apparent contours in a sequence of images have already been studied by several a~thors.~,~~~~' Most of the proposed curved surface reconstruction methods require an accurate estimation of the camera motion, and sometimes use a camera model more simple than the standard perspective projection. Moreover, in the case of solids of revolution, no approach has taken advantage, as far as we know, of a left-right limb matching to improve the estimation perforniances. The proposed method consists in analyzing the dynamic of apparent contours in a sequence of monocular images. It is based on the discrete feature paradigm and on a cylindrical model to represent objects in a static scene. It enables the simultaneous estimation of the camera motion and cylindrical object parameters. In this work, the camera is considered to be embedded in a robotic platform, the control system of which provides a camera motion a prioriestimation with an error up to 20 percents. No assumptions are made concerning the cylindrical scene: number of cylinders, radii,. . .The matching of apparent contours in the sequence is performed by a visual module not considered herein. The dynamic and nonlinear features of the system, as well as the computation time burden led us to use the nonlinear Kalman filter (socalled extended Kalman filter). This research can be related to the work of Broida el al.,' who presented a method for estimating the kinematics and structure of a rigid object from a long sequence of images. 3 MODELS