Detection by registration

Abstract

The Detection by Registration (DbR) method is proposed as an efficient and accurate solution to the bin picking problem. DbR exploits the relatively large convergence basin of the Iterative Closest Point (ICP) algorithm, by specifying a set of characteristic poses from which to initiate ICP for a given part. For each characteristic pose, an efficient search is executed at a set of uniform seed point translations in a scene, by incrementing the pose through a small set of discrete in-plane rotations. The results from this search are then refined further using ICP and validated. Experiments have shown favorable performance, with an average of 600% detection accuracy and an average execution time of less than 400 msec per scene.