3-D motion and shape from multiple image sequences

We propose a method for determining the accurate local 3-D motion and shape of objects in image sequences from multiple views. The shape of the considered object is calculated at each time step using image sequences, synchronously acquired by two or more calibrated cameras. Starting from an initial shape model of the object a stepwise refinement process is performed in which the model is adapted to the actual shape according to a cost function. This function maximizes the similarity between the respective model projections of the object onto the images. In order to obtain an appropriate initial shape model at the consecutive time step a motion estimation technique for monocular image sequences is applied. The proposed method is applicable for rigid and deformable objects as well. Selected results for real-world image sequences are presented.