6D Object Pose Estimation Algorithm Using Preprocessing of Segmentation and Keypoint Extraction

Abstract

In this paper we will explain a 6D (x, y, z, yaw, pitch, roll) object pose estimation algorithm. The point cloud based template matching method is used to solve this problem. To reduce the complexity of this task, two preprocessing methods are performed. The first method of preprocessing is a 2D/2.5D object detector for scene point cloud segmentation. The second is a point cloud based keypoint extractor which is used for the template point cloud. In order to improve the performance of the preprocessing, an encoder is used. By using the encoder three geometric features (local point density, difference of normal and curvature) are extracted. The features and raw data (RGB-D and point cloud) are combined into the encoded multimodal data. Finally, the experiments show the improvement of object pose estimation. Meanwhile in contrast to difference of normal and curvature, local point density is the optimal feature to perform scene point cloud segmentation.