Efficient and Distributed Large-Scale 3D Map Registration using Tomographic Features

TL;DR

The paper tackles large-scale, distributed 3D map registration by introducing tomographic features extracted from gravity-aligned maps as horizontal slices. A two-stage, consensus-based registration aligns 4 DoF transforms by first solving a 3 DoF 2D problem at fixed height differences and then selecting the height offset to complete the transformation, leveraging strong parallelism. Across simulated and real datasets, the tomographic approach (especially Consensus and Tomographic-TEASER++) demonstrates superior memory efficiency and competitive or superior accuracy, with notable resilience to measurement noise compared to learning-based baselines. This method enables scalable map merging for multi-robot systems on resource-constrained hardware, with clear pathways to future non-rigid extensions and roll/pitch recovery under gravity alignment assumptions.

Abstract

A robust, resource-efficient, distributed, and minimally parameterized 3D map matching and merging algorithm is proposed. The suggested algorithm utilizes tomographic features from 2D projections of horizontal cross-sections of gravity-aligned local maps, and matches these projection slices at all possible height differences, enabling the estimation of four degrees of freedom in an efficient and parallelizable manner. The advocated algorithm improves state-of-the-art feature extraction and registration pipelines by an order of magnitude in memory use and execution time. Experimental studies are offered to investigate the efficiency of this 3D map merging scheme.

Figures (15)

• Figure 1: Samples of 2D binary slice images extracted at differnet heights for a simulated indoor environment. Binary images with colorized borders denote the corresponding height at the 3D rendering of the environment.
• Figure 2: A visual summary of the proposed distributed 3D map matching framework. Each agent $\{c, d\}$ is responsible for extracting slices ${^{\{c, d\}}}\mathbf{s}_{h_i}$ at a predefined grid size. The pose is estimated via two-stage optimization algorithm, where one of the agents cross-correlates the slices by estimating a 3DoF rigid transformation, $\widetilde{\mathbf{T}}_d^c$, between slices, and the consensus of different height hypotheses, $z_d^c$, is used to determine the 4th DoF.
• Figure 3: A sample pair of color and depth images from the InteriorNet dataset.
• Figure 4: Sample point clouds from InteriorNet dataset under various noise conditions.
• Figure 5: Translation errors of the tested algorithms for the InteriorNet data on various noise levels.