A Fusion of CNNs and ICP for 3-D Point Cloud Registration

3-D point cloud registration appears to be one of the principal techniques to estimate object pose in 3-D space and is critical to object picking and assembly in automated manufacturing lines. Thereby, this paper proposes an effective registration architecture with the aim of estimating the transformation between a data point cloud and the model point cloud. Specifically, in the first registration stage, a trainable Convolutional Neural Network (CNN) model is developed to learn the pose estimation between two point clouds in the case of a full range of orientation from −180° to 180°. In order to generate the training data set, a descriptor is proposed to extract features which are employed to train the CNN model from point clouds. Then, based on the rough estimation of the trained CNN model in the first stage, two point clouds can be further aligned precisely in the second stage by using the Iterative Closest Point (ICP) algorithm. Finally, the performance of the proposed two-stage registration architecture has been verified by experiments in comparison with a baseline method. The experimental results illustrate that the developed algorithm can guarantee high precision while significantly reducing the estimation time.