Pose Determination By PotentialWell Space Embedding

Abstract

A novel algorithm is introduced to estimate the pose of objects from sparse range data. Pose determination is tackled by employing the ICP algorithm to find corresponding local minima between a preprocessed model and the runtime data. Unlike other existing algorithms that try to avoid local minima, here local minima are used as effective feature vectors for generating multiple hypotheses of the pose. These hypotheses are then examined and verified using the bounded Hough transform, which is more robust than using the registration error directly. Only a small number of iterations (e.g., 5) is needed for each ICP at both preprocessing and runtime, which makes the technique efficient. The algorithm has been implemented and tested on a variety of objects, including freeform models, using both simulated and real data from Lidar and stereovision sensors. The experimental results show the technique to be both effective and efficient, executing at multiple frames per second on standard hardware. In addition, it functions well with very sparse data, possibly comprising only hundreds of points per frame, and it is also robust to measurement error and outliers.