A Multi-Robot 3D Point Cloud Map Merging Method Based on Overlapping Region

In multi-robot mapping, differences in sensor capabilities across various robotic platforms can result in significant discrepancies in the generated maps. A more critical issue is the large displacement between maps, making it difficult to determine the transformation relationships between them accurately. Additionally, multi-robot SLAM often requires processing a larger number of points, leading to substantial computational resource consumption. This paper presents a two-stage alignment algorithm for merging 3D point cloud maps. The algorithm does not require any initial guesses for the transformations between local maps but instead relies on the overlapping regions of the maps to align and merge them. Initially, the original maps are segmented and sub-maps are extracted to reduce the number of points, thereby accelerating the subsequent processing efficiency. The extracted sub-maps are then globally aligned with the local maps from another robot, resulting in a rough transformation matrix. This global alignment solution is used as the initial guess for point-based fine registration, subsequently refined. In this process, we use a voxelized Generalized Iterative Closest Point fine registration algorithm to avoid expensive nearest neighbor searches, thereby increasing registration speed while maintaining accuracy. Finally, the data is merged to obtain an accurate global map. The efficiency and accuracy of the proposed 3D point cloud map merging method have been validated through experiments on public datasets.

I. Introduction

When multiple robots participate in the SLAM process, each robot generates a local map within its local frame. These local maps serve as sources of information for localization, obstacle avoidance, navigation, and path planning. Later, these maps can be shared and merged into a global map to provide a better representation of the environment, thereby enabling efficient and accurate task completion. In practical applications, the complexity of the map merging process is determined by various factors. These include the knowledge of the relative positions and orientations of the robots, as well as the correct fusion of sensor data from each robot. The sensors may have different accuracies, inherent noise, or ranges. The differences in sensors across different robotic platforms (aerial or ground) can lead to significant discrepancies in the generated maps, making direct merging of these maps more challenging [1] . A feasible solution is to rely on the overlapping regions between the local maps generated by each robot to align and merge the maps [2] . The most challenging and critical part of this process is to find the transformation relationship between the overlapping regions (point cloud registration).