Discrete Pose Space Estimation to Improve ICP-Based Tracking

Abstract

Iterative closest point (ICP)-based tracking works well when the interframe motion is within the ICP minimum well space. For large interframe motions resulting from a limited sensor acquisition rate relative to the speed of the object motion, it suffers from slow convergence and a tendency to be stalled by local minima. A novel method is proposed to improve the performance of ICP-based tracking. The method is based upon the bounded Hough transform (BHT) which estimates the object pose in a coarse discrete pose space. Given an initial pose estimate, and assuming that the interframe motion is bounded in all 6 pose dimensions, the BHT estimates the current frame's pose. On its own, the BHT is able to track an object's pose in sparse range data both efficiently and reliably, albeit with a limited precision. Experiments on both simulated and real data show the BHT to be more efficient than a number of variants of the ICP for a similar degree of reliability. A hybrid method has also been implemented wherein at each frame the BHT is followed by a few ICP iterations. This hybrid method is more efficient than the ICP, and is more reliable than either the BHT or ICP separately.