End to End Robust Point-Cloud Alignment Using Unsupervised Deep Learning

Abstract

The point-cloud alignment methods help robots to map their environment, recognize target objects and estimate rigid-body object poses from the 3D vision sensor data. In this paper, we propose a robust and computationally efficient approach for point-cloud alignment. Unlike the feature descriptor-based pose classifiers or regression methods, the proposed method can process an unordered point cloud by mapping it uniquely onto a particular 2D space determined based on the point cloud from the object. The model training is fully unsupervised and relies on optimizing the projection results based on a loss function. Specifically, the proposed 2D mapping enables the model to recognize objects with a simple linear classifier to increase computational efficiency. Then, the proposed method calculates the object pose in the continuous space rather than classifying the point cloud into discrete pose labels. The experiments and comparison with a well-established descriptor-based point-cloud alignment method show that the proposed method has a good performance and is robust to missing points of the point cloud. The higher performance in recognition and pose estimation precision make the method suitable for industrial robotic and automation applications.