Registering 3D point clouds with low overlap is challenging in 3D computer vision, primarily due to difficulties in identifying small overlap regions and removing correspondence outliers. We observe that the neighborhood similarity can be utilized to detect point correspondence, and the consistent neighborhood correspondence can be used as a criterion to detect robust overlapping regions. So that a Double-layer Multi-scale Star-graph (DMS) structure is proposed to detect robust correspondences using two different types of multi-scale star-graphs. The first-layer Multi-scale Neighbor Feature Star-graphs (MNFS) takes each point as the center and its multi-scale nearest neighbors as the leaves. The MNFS enables to establish the initial correspondence candidate set between the two point clouds based on multi-scale neighborhood topology and feature similarity. Subsequently, each pair of corresponding points find their nearest neighbors within the correspondence sets to construct a Multi-scale Matching Star-graphs (MMS) on each side, so the mutual correspondence relationships between the MMS vertices are identified. These identified mutual correspondences are treated as vertices to construct the Multi-scale Correspondence Star-graphs (MCS), that indicate the relationships among the correspondences. We design edge weight and vertex weight criterion in MCS to detect only the robust correspondence set that has strong neighborhood consistency, so as to reject the outliers. Finally, the point cloud registration is conducted based on the detected robust correspondence. The experimental results demonstrate clearly that the proposed DMS method exhibits superior robustness when compared to existing state-of-the-art registration algorithms.

Registering multiple point clouds (MPC) to form a consistent global map is a crucial problem for various 3D mapping applications, which is however highly challenging due to the tightly coupled difficulties of: 1) identifying the registrable point cloud pairs (RPCPs), 2) detecting the point correspondence in each RPCP, 3) estimating the relative transformation for each RPCP, and 4) aligning the RPCPs to generate the global map. This paper proposes an Iterative Pairing and Transformation (IPT) framework to address above difficulties. A Registrability Graph (RG) is in particular proposed, whose vertices represent the $n$n point clouds and its edges model the registerability between each pair of the point clouds. Registerability is an indicator for whether the relative transformation of the two point clouds can be trustfully estimated. Starting from an $N$N-clique, the obviously unregistrable pairs will be identified through an adaptive coarse registration stage using RANSAC, leading to a sparser RG. For the remained edges in RG, the relative transformation for each registrable pair is accessed by a learning-based pairwise registration algorithm. These pairwise relative poses among vertices are then input into the global registration step, which optimizes and generates globally consistent pose for each point cloud, as well as a global map. Then, based on the global poses of the point clouds, the edges of the RG will be reevaluated to adjusting the registrable pairs, and this local pairing and global optimization process will repeat until convergence. The experimental results show IPT outperforms current state-of-the-art MPC registration algorithms.

In the work, we present a novel multiviewpoint and multitemporal remote sensing image registration method based on a dual-features Student-$t$distribution mixture model (DSMM) under a variational Bayesian (VB) framework. The main contributions of the work are: 1) guided image filter (GIF) is adopted to smooth edges and strengthen ridges of images for heightening characteristics of feature point; 2) a Student-$t$distribution mixture model (SMM) based DSMM designs a global and local descriptor to estimate correspondences from local to global scale; and 3) local structure constraints are designed to preserve relationships of neighbors of points and the scale of neighborhood structure of points to constrain transformation. The experimental results demonstrate the better performance of our DSMM against five state-of-the-art methods.

Natural disasters frequently have caused huge impact on life and property losses, in Southwest China. To provide assistance for disaster relief, areas damaged in natural disasters is quickly located by utilizing satellite remote sensing images based-deep learning object detection technology. However, the current detection technology, for the detection of damaged objects discretely in the disaster area, has some challenges, such as partial missing of multi-source images and extremely sparse targets with weak features or occlusion at large scales. Furthermore, we propose an object detection network based on dynamic extraction of multi-source images features to solve above problems. To train our proposed network, we collect multi-source remote sensing images before and after the disaster. Finally, it is verified that when the detection error rate is less than 5%, the accuracy of the detection model reaches more than 85%.

Point-set registration is widely used in computer vision and pattern recognition. However, it has become a challenging problem since the current registration algorithms suffer from the complexities of the point-set distributions. To solve this problem, we propose a robust registration algorithm based on the estimation of distribution algorithm (EDA) to optimize the complex distributions from a global search mechanism. We propose an EDA probability model based on the asymmetric generalized Gaussian mixture model, which describes the area in the solution space as comprehensively as possible and constructs a probability model of complex distribution points, especially for missing and outliers. We propose a transformation and a Gaussian evolution strategy in the selection mechanism of EDA to process the deformation, rotation, and denoising of selected dominant individuals. Considering the complexity of the model, we choose to optimize from the perspective of variational Bayesian, and introduce a prior probability distribution through local variation to reinforce the convergence of the algorithm in dealing with complex point sets. In addition, a local search mechanism based on the simulated annealing algorithm is added to realize the coarse-to-fine registration. Experimental results show that our method has the best robustness compared with the state-of-the-art registration algorithms.