3D motion estimation from a stereo image sequence using dual-sequential-Kalman-filter

Aiming at solving the coupling and time-consuming problem in motion estimation from images, a recursive estimator comprised of two sequential Kalman filters is proposed. 3D motion of a rigid object can be decomposed into translation of a point fixed on the object, called rotation center, and rotation w.r.t. this point. The rotational parameters are proved to be separate with the others, which means the motion has the potential to be decoupled. Viewing the moving object as a dynamic system, called moving object system, motion estimating is formulated as a state estimation problem. Decoupling the moving object system into two sub-systems, then the dual-sequential-Kalman-filter can be designed to estimate the states of the moving object system, thus a high dimension filter is replaced with two reduced ones. As time cost in computing depends on the third power of the dimension of the estimator, the time-consuming problem is solved partly. The performance of dual-sequential-Kalman-filter is illustrated using both simulated and real image sequences, two important merits, accuracy and robustness, are presented with the experiment results.