Scan registration for mechanical scanning imaging sonar using kD2D-NDT

• Main Authors: Jiang, Min, Song, Sanming, Li, Yiping, Liu, Jian, Feng, Xisheng
• Format: Conference Proceeding
• Language: English
• Subjects: Device-to-device communication; Euclidean distance; Gaussian distribution; Gaussian mixture model; Imaging; kD2D-NDT; mechanical scanning imaging sonar; scan registration; Sonar; Three-dimensional displays
• ISSN: 1948-9447
• DOI: 10.1109/CCDC.2018.8408259

A method derived from the D2D-NDT, named kD2D-NDT, is proposed to register the scans that are collected by the Mechanical Scanning Imaging Sonar (MSIS). The D2D-NDT method replaces the point-to-distribution (P2D) scoring in the normal distribution transformation (NDT) with distribution-to-distribution (D2D) matching, greatly reducing the computation cost. In this paper, several heuristic strategies are adopted in kD2D-NDT to accelerate and stabilize the matching process. Firstly, the point cloud of the floating scan and the reference scan are grouped into compact clusters by the K-means clustering method to accommodate the Gaussian mixture model assumption which underlies the D2D distance measure and no iterative optimization at different grid size is needed. Secondly, for each Gaussian component in the floating scan, only k =3D 3 nearest Gaussian components in the reference scan are chosen to measure the similarity. Lastly, to avoid the singularity in calculating the matrix inverse, the Euclidean distance between the centroid pair, instead of the Mahalanobis distance, is adopted to find the most similar Gaussian components. Its applications to the scans that are collected from the realistic underwater environment show that the proposed strategies make kD2D-NDT practical for the MSIS scans.