Alinhamento espaço-temporal de sequências de vídeo capturadas a partir de múltiplos pontos de vista

Abstract

This thesis addresses the problem of estimating the spatio-temporal alignment between N unsynchronized video sequences of the same dynamic 3D scene, captured from distinct viewpoints. Unlike existing methods, which work for N = 2 and rely on a computationally-intensive search in the space of temporal alignments, we present a novel approach that reduces the problem for general N to the robust estimation of a single line in RN. This line captures all temporal relations between the sequences and can be computed without any prior knowledge of these relations. Considering that theparameters of fundamental matrices capture the spatial alignment, we use an iterative algorithm to refine simultaneously the parameters representing the temporal and spatial relations between the sequences, since that the exclusive refinement of the temporal parameters is suboptimal. Experimental results with real-world sequences show that our method can accurately alignvideos even when they have large misalignments (e.g., hundreds of frames), when the problem is seemingly ambiguous (e.g., scenes with roughly periodic motion), and when accurate manual alignment is dicult (e.g., due to slow-moving objects). Finally, experiments with synthetic sequences demonstrate the scalability and accuracy of our approach, providing quantitative measurements for the quality of the spatio-temporal alignments estimated.