f \ AN ACTIVE ROBOT VISION SYSTEM FOR REAL-TIME 3-D STRUCTURE RECOVERY S. BOUKIR*. F. CHAUMETTE*, P. BOUTHEMY* and D. JUVIN** ' IRISA / INRIA, Campus de Beaulieu, 35042 Rennes Cedex, France. " CEA-LETI / DEIN-SLA Saclay, 91191 Gif sur Yvette, France. Abstract. This paper presents an active approach for the task of computing the 3-D structure of a nuclear plant environment from on image sequence, more precisely the recovery of the 3-D structure of cylindrical ob- jects. Active vision is considered by computing adequate camera motions using image-based control laws. This approach requires a real-lime tracking of the limbs of the cylinders. Therefore, an original matching approach, which relies on an algorithm for determining moving edges, is proposed. This method is distinguished by its robustness and its easiness to implement. This method has been implemented on a parallel image processing board and real-time performance has been achieved. The whole scheme iias been successfully validated in an experimental set-up. Key words. Closed-loop systems; image processing; least-squares estimation; nuclear plants; optimal es- timation: parallel processing; parameter estimation; real-lime computer systems; target tracking. 1. INTRODUCTION The recovery of 3-D descriptions of a scene from images is one of the main issues in computer vi- sion. The work investigated here is concerned with the use of sequences of images acquired by a moving camera mounted on the end effector of a robot arm to get an exact and complete descrip- tion of the environment. This non-trivial issue requires the development of efficient algorithms dealing with the analysis and the interpretation of dynamic visual information. The output of this processing step could be used for inspection tasks. The major shortcomings that often limit the performance of current systems are their sen- sitivity to noise and their unsatisfactory accuracy. An attractive way to cope with these problems is to follow an active vision approach (Aloimonos et al., 1987; Bajcsy, 1988) which can be defined as an intelligent data acquisition process. Therefore, the problem is to elaborate control strategies for adaptively setting sensor parameters, in order to improve the knowledge of the environment. This paper presents an application of the concept of active vision for 3-D structure recovery in the context of nuclear power plants which are gener- ally encumbered with numerous pipes. The prob- lem has been examined under three main aspects : • to find reliable pairings of contour segments in two successive images by extending an al- gorithm which determines moving edges in an image sequence (Bouthemy. 1989), • so define an optim.il method for 3-D recon- struction of geometric features ( Boukir and Cliaiiinette. 1092); this method is here ap- plied to the particular case of cylindrical prim- itives which are an appropriate model for pipes, and • to effectively implement the whole scheme (i.e., vision and control) on an experimen- tal set-up taking into account real-time con- straints. 2. A LOCAL APPROACH FOR MATCHING CONTOUR SEGMENTS Matching features extracted from two images is a basic low-level issue in machine vision and is a key step in structure from motion. The task of estab- lishing and maintaining such correspondences re- mains complex. Part of the difficulty comes from discontinuities and occlusions that occur in real- world scenes. Since the application aimed at here is the reconstruction of cylindrical objects from their successive projections in the image sequence, contour segment seems to be a natural choice for an image primitive. Many classes of methods have been developed for the matching of such primi- tives. A commonly used method is to introduce a temporal recursive filter to track contour segments (Deriche, 1990; Crowley, 1992), and to introduce a prediction step. Other methods rely on contextual information to alleviate possible matching ambi- guities (Medioni, 1984; Chen and Huang, 1990). The matching is performed at the level of contour segments determined by a linear approximation of edges. Then, the available descriptors are global properties such as length and orientation. These techniques are obviously sensitive to segmentation instability and to occlusion. Furthermore, such methods may become too complex to be imple- mented for real-time performance and will require dedicated boards.