3-D SCENE RECONSTRUCTION FROM MONOCULAR IMAGE SEQUENCES

Milan Hanajík - Paul P. J. van den Bosch

   In this paper we deal with the problem of 3-D scene description reconstruction from monocular image sequences. The reconstruction is based on the processing of linear features extracted from the acquired images. Two reconstruction algorithms are introduced. The first one formulates the problem as the stochastic filtering problem and the algorithm computes the scene description incrementally using the Extended Kalman Filter (EKF). Issues of the computational complexity of the algorithm are addressed and an algorithm with computation time in each iteration step linearly proportional to the number of processed line segments is proposed. The second algorithm formulates the problem as the global optimization problem, and the maximum likelihood estimate (MLE) of the scene is numerically computed from all images in one batch. Results achieved with a sample image sequence are given and conclusions are derived.

Keywords: 3-D structure from motion, Extended Kalman Filter, computational complexity

[full-paper]