Intelligent Systems Laboratory

Robust Pose Estimation

Once a set of more than three points of a three--dimensional object have been matched to a corresponding set of image points, the pose of the object (its position and orientation) can be determined by solving a non--linear set of equations. Should it happen, as it occasionally does, that not all of the correspondences between the two--dimensional and three--dimensional points are correct, then the incorrect correspondence(s) can have an inordinately significant effect on the solution of the non--linear equation. Indeed, under these conditions, the determined solution may be very far from the correct solution. There is a methodology of solving the non-- linear equation for the estimated pose which, under the conditions of a few incorrectly matched points, will produce a solution which is almost as good as if the only points given it were those in correct correspondence. This methodology is called robust estimation. Using this methodology we are developing a technique which can tolerate a few mismatched points and compute reasonably good estimates of the correct pose. The efficacy of the technique has so proved itself in tens of thousands of controlled experiments.