Mise en correspondance active et passive pour la vision par ordinateur multivue

Papyrus

http://hdl.handle.net/1866/17860

Abstract

This thesis explores an important aspect of computer vision named multi-view matching. A view can either be a passive device such as a camera or an active one such as a multimedia projector. Our goal is to find the pixels in different views that correspond to the same 3D point of a scene. Many new energy-based methods for active and passive multi-view matching are presented. The thesis is composed of two main parts. Whilst the first one presents novel passive algorithms, the second describes a new algorithm for active matching. In the first part, three new passive multi-camera matching algorithms are introduced. These are designed to overcome the occlusion problems of multi-baseline stereo. The first one detects occlusions in the depth map obtained from regular efficient stereo matching algorithms. Occlusions are detected as inconsistencies of the depth map by computing the visibility of the map as it is reprojected into each camera. The matching cost function is modified according to the detected occlusions by removing the offending cameras from the computation of the matching cost. The algorithm gradually modifies the matching cost function according to the history of inconsistencies in the depth map, until con vergence. We also provide an extension to this framework that features a non-uniform spatial decomposition of the disparity map. This extension allows the integration of the visibility computation and refinement of the disparity into a single iterative frame work. While two graph-theoretic stereo algorithms are used in our experiments, our framework is general enough to be applied to many others. The second algorithm is a hybrid between fast heuristic occlusion overcoming algorithms that precompute an approximate visibility and slower methods that use correct visibility handling. Our approach is based on iterative dynamic programming and computes simultaneously C disparity and camera visibility. Interestingly, dynamic programming makes it possi ble to compute exactly part of the visibility information. The remainder is obtained through heuristics. Finally, we also propose an algorithm that improves the Iocaliza tioll of disparity discontinuities of disparity maps obtained by multi-baseline stereo matcher. Rather than associating a disparity label to every pixel of a disparity map, it associates a position to every disparity discontinuity. This formulation allows us to find an approximate solution to a 2D labeliiig problem with robust smoothing terrn by minimizing multiple 1D problems, thus making possible the use of dynamic program ming. The validity of the proposed novel algorithms is established using real imagery with ground truth and they compare favorably with other state-of-the-art algorithms. In the second part of the thesis, an active matching algorithm is introduced. It is a new real-time energy-based formulation of the matching problem encountered when establishing the correspondence using a stripe-based structured light system. We pro- vide a general framework that takes into consideration the projective displacement which is induced by the camera, projector and scene configuration. This allows the selection of a rieighborhood for the smoothing term which is geometrically plausible. The minimization uses a reformulation of the dynamic programming recurrences and is performed on a programmable graphic processor. We show experimentally that our approach provides major improvements under difficult conditions. Our method is also used for projector calibration: we validate the results with a laser tracker.

Documents

PDF

---