Limited memory filter for 3D relative position orientation estimation of maneuvering target

In a vision-based positioning system, noise in the image features results in errors in 3D reconstruction, and thereby, causes a serious degradation in individual 3D object pose (position and orientation) estimation. To find a solution for this problem, a filter approach is presented for more accurate relative 3D pose estimation of a moving object with respect to the vision sensor system. Different from previous schemes that were limited to the assumption that the relative motion between the target and the camera is smooth and slow, this approach is implementable for a maneuvering target without a prior knowledge about its dynamics. And the computation load is much reduced by avoiding the need of EFK. This filter is composed of a maneuver detector and a filter updater. First, by analyzing the effect of image noise on the pose estimate error, linear measure equations relating the nominal pose parameters (obtained by 3D reconstruction) and the true ones are proposed. Then, a maneuver detector of maximum detection probability is deduced, and two sub-optimal detectors for quick maneuver and slow maneuver are further developed respectively. When a maneuver is declared, updating is performed with limited memory filtering. Experiment results verify the effectiveness of this approach