loading page

An Efficient Large-Scale 3D Map Stitching Algorithm using Automatic Overlapping Area Identification
  • +1
  • Hsien-I Lin,
  • Muhammad Ahsan Fatwaddin Shodiq,
  • An-Kai Jeng,
  • Chun-Wei Chang
Hsien-I Lin
National Yang Ming Chiao Tung University

Corresponding Author:[email protected]

Author Profile
Muhammad Ahsan Fatwaddin Shodiq
National Yang Ming Chiao Tung University
Author Profile
An-Kai Jeng
Industrial Technology Research Institute
Author Profile
Chun-Wei Chang
Industrial Technology Research Institute
Author Profile

Abstract

Quality of 3D point cloud maps is essential for navigation and localization in Autonomous Mobile Robots, yet creating these maps for large-scale areas presents challenges, stemming from the processing of numerous points. In such situations, constructing a 3D map can be accomplished by dividing it into smaller regions and then merging them to generate a complete map by performing a 3D map stitching algorithm. Currently, these overlapping areas are manually selected, which leads to potential errors. In response, a novel method to automatically identify overlapping areas is proposed to perform map stitching based on the overlapping areas only instead of the entire maps. Utilizing the proposed method results in a significant reduction in time consumption. The proposed automatic method incorporates the DBSCAN algorithm for area clustering, template matching for correspondence cluster identification, and a binary-search algorithm for determining the best parameter. For the performance evaluation, the proposed method was compared with manual selection and with the use of the entire maps in the map-merge-3D algorithm. The method achieves a significant reduction in the time required for the 3D map stitching process, amounting to a 38.64% decrease compared to using the entire maps. In terms of accuracy, the proposed method reduces translation error to 0.1723m and rotation error to 0.1763 , representing decreases of 5.28% and 16.16%, respectively, while manual selection results in a translation error of 0.4278m and rotation error of 0.7123 , increases of 135.25% and 238.75% respectively, compared to the entire maps 0.1819m and 0.2103 .