Surface reconstruction for multi-view video

This thesis introduces a methodology for obtaining an alternative representation of video sequences captured by calibrated multi-camera systems in controlled environments with known scene background. This representation consists in a 3D description of the surfaces of foreground objects, which allows for the recovering of part of the 3D information of the original scene lost in the projection process in each camera. The choice of the type of representation and the design of the reconstruction techniques are driven by three requirements that appear in smart rooms or recording studios. In these scenarios, video sequences captured by a multi-camera rig are used both for analysis applications and interactive visualization methods. The requirements are: the reconstruction method must be fast in order to be usable in interactive applications, the surface representation must provide a compression of the multi-view data redundancies and this representation must also provide all the relevant information to be used for analysis applications as well as for free-viewpoint video. Once foreground and background are segregated for each view, the reconstruction process is divided in two stages. The first one obtains a sampling of the foreground surfaces (including orientation and texture), whereas the second provides closed, continuous surfaces from the samples, through interpolation. The sampling process is interpreted as a search for 3D positions that result in feature matchings between different views. This search process can be driven by different mechanisms: an image-based approach, another one based on the deformation of a surface from frame to frame or a statistical sampling approach where samples are searched around the positions of other detected samples, which is the fastest and easiest to parallelize of the three approaches. A meshing algorithm is also presented, which allows for the interpolation of surfaces between samples. Starting by an initial triangle, which connects three points coherently oriented, an iterative expansion of the surface over the complete set of samples takes place. The proposed method presents a very accurate reconstruction and results in a correct topology. Furthermore, it is fast enough to be used interactively. The presented methodology for surface reconstruction permits obtaining a fast, compressed and complete representation of foreground elements in multi-view video, as reflected by the experimental results. Aquesta tesi presenta di