3-D Point Cloud Registration Using Convolutional Neural Networks

Abstract

This paper develops a registration architecture for the purpose of estimating relative pose including the rotation and the translation of an object in terms of a model in 3-D space based on 3-D point clouds captured by a 3-D camera. Particularly, this paper addresses the time-consuming problem of 3-D point cloud registration which is essential for the closed-loop industrial automated assembly systems that demand fixed time for accurate pose estimation. Firstly, two different descriptors are developed in order to extract coarse and detailed features of these point cloud data sets for the purpose of creating training data sets according to diversified orientations. Secondly, in order to guarantee fast pose estimation in fixed time, a seemingly novel registration architecture by employing two consecutive convolutional neural network (CNN) models is proposed. After training, the proposed CNN architecture can estimate the rotation between the model point cloud and a data point cloud, followed by the translation estimation based on computing average values. By covering a smaller range of uncertainty of the orientation compared with a full range of uncertainty covered by the first CNN model, the second CNN model can precisely estimate the orientation of the 3-D point cloud. Finally, the performance of the algorithm proposed in this paper has been validated by experiments in comparison with baseline methods. Based on these results, the proposed algorithm significantly reduces the estimation time while maintaining high precision.

Building a Next Generation AI Platform for AEC: A Review and Research Challenges

“…When using the ModelNet40 database to test, the DCP algorithm is robust when facing the noise, and its accuracy is higher than PointNetLK (Wang and Solomon 2019). Another method used 3D-CNN and 2D-CNN for coarse registration and fine registration respectively (Chang and Pham 2019). For accuracy, this method is close to RANSAC + ICP algorithm, and its efficiency is 15 times that of RANSAC + ICP algorithm.…”

When Deep Learning Meets Data Alignment: A Review on Deep Registration Networks (DRNs)

“…Chang and Pham [63] presented a 3D point set registration framework with two stages to cover the problem of coarse-to-fine registration. Two descriptors are proposed, one for rough and one for fine orientation extraction, SSPD and 8CBCP, respectively.…”

“…• Registration of global and local point clouds with a generic Deep Auto Encoder (DAE) architecture regardless of the input data and its source [15] • The Shape Deformation Network (SDN) is an autoencoder architecture able to deal with deformable shapes extracting global shape descriptors [29] • Capability for multi-spectral registration using Asymmetric Siamese Convolutional Networks [57] • Filtering inaccurate features correspondences based on geometrical and global properties to minimize their influence in the registration process [67,73] • Registration process speedup using a two-staged approach based on Convolutional Neural Networks (CNNs) to solve coarse-to-fine registration problems [63] transform • Using Spatial Transformer Networks (STNs) to deal with geometrical variability by means of learned invariance to transformations [45,47,95] • To align multimodal inputs with transformations generated and evaluated in an adversarial fashion [60] • Ability to preserve detailed geometric information by using a CNN to infer Free Form Deformation (FFD) parameters for 3D template-image matching [65] • Unsupervised registration of multiple point clouds in a global frame of reference [27] • Outperforming single channel CNNs with a multi-channel approach for real-time deformable registration [64] • Cloth fitting by modeling 3D body-garment interaction in real time surpassing Physics-Based Simulation (PBS) [36]…”

DGCB-Net: Dynamic Graph Convolutional Broad Network for 3D Object Recognition in Point Cloud

“…Different from the deep learning models that employed the gradient descent method on the training process, the parameter optimizing and fine-tuning processes of the broad structure were faster and more concise. In the feature layer, a series of mapping matrix W j were randomly initialized to project X to m feature layers Z j according to Equation (5). Different from the relationships among convolutional layers, feature layer Z j was not influenced by the last layer Z j−1 in broad networks.…”

“…The inherent characteristics mainly include uncertain vertex topology, uneven point density, arbitrary point count, and unfixed point permutation [4]. To avoid the influence caused by these characteristics, point cloud pre-processes, or registration steps are required in the majority of 3D object recognition methods [5].…”