Vision-based motion tracking of frigid objects using prediction of uncertainties

A vision-based motion tracking method described in this paper estimates the 3D position and orientation of a moving object of known shape at an average speed of 2.5 seconds per image frame even in complex environments using a conventional computer power. Given a coarse estimate of the initial 3D object pose, the method first generates apt expectation view from which visible model features are automatically selected. The method then extracts potentially matched image features from image regions bounded by the propagation of object motion uncertainty. The special aspect of our vision-based trading is an optimal correspondence search for model features and image features in which we use a Kalman filter-based updating scheme to perform the precise 3D object pose estimation. Experimental results are presented to demonstrate the robustness of the method even in the presence of occlusion