Recognition, Pose and Tracking of Modelled Polyhedral Objects by Multi-Ocular Vision

Abstract

We developped a fully automated algorithmic chain for the tracking of polyhedral objects with no manual intervention. It uses a multi-cameras calibrated system and the 3D model of the observed object. The initial phase of the tracking is done according to an automatic location process using graph theoretical methods. The originality of the approach resides mainly in the fact that compound structures (triple junction and planar faces with four vertices) are used to construct the graphs describing scene and model. The association graph construction and the search of maximal cliques are greatly simplified in this way. The final solution is selected among the maximal cliques by a predictionverification scheme. During the tracking process, it is noticeable that our model based approach does not use triangulation although the basis of the multi-ocular system is available. The knowledge of calibration parameters (extrinsic as well as intrinsic) of the cameras enables to express the various equations related to each images shot in one common reference system. The aim of this paper is to prove that model based methods are not bound to monocular schemes but can be used in various multi-ocular situations in which they can improve the overall robustness.