Adaptive Pose Estimation Algorithm Based on Point Pair Feature

Abstract

Accurately estimating the 6-D pose is crucial in various modern production fields. This article introduces a novel approach for efficiently and precisely estimating the 6-D pose of known objects within point cloud scenes of various resolutions, aiming to tackle the challenge of rapid and accurate 6-D pose estimation. This framework utilizes point pair features (PPFs) to conduct pose voting through a combination of offline training and online matching. Scene coefficients and model coefficients are introduced, and the required neighborhood point number for filtering is calculated based on different target objects, scenes, and point cloud densities. The appropriate downsampling rate is also calculated, and a novel hypothesis verification method is utilized to eliminate false positives and predict the accurate 6-D pose of the target object. The framework demonstrates good performance on industrial part datasets, showing good robustness under different point cloud densities.