Large-scale DSM registration via motion averaging

TL;DR

This work tackles large-scale DSM registration by reframing the problem as motion averaging over a scene graph built from all pairwise DSM registrations. It introduces DSM-ICP, a grid-based, memory-efficient pairwise registration method, enabling exact nearest-neighbor searches without caching the entire reference dataset. A weighted scene graph is then constructed and optimized via a motion-averaging approach to yield global poses, decoupling rotation ($R_i$ via SVD) from translation ($t_i$ via least squares) and achieving $O(N)$ complexity for the global optimization. Empirical results on public satellite DSMs with lidar ground truth show substantial improvements in both computation and accuracy over k-d tree-based ICP and greedy MST approaches, including reduced accumulated error and better reconstruction quality. The method offers scalable, accurate large-area DSM registration suitable for integration into geospatial pipelines and large-scale mapping workflows.

Abstract

Generating wide-area digital surface models (DSMs) requires registering a large number of individual, and partially overlapped DSMs. This presents a challenging problem for a typical registration algorithm, since when a large number of observations from these multiple DSMs are considered, it may easily cause memory overflow. Sequential registration algorithms, although can significantly reduce the computation, are especially vulnerable for small overlapped pairs, leading to a large error accumulation. In this work, we propose a novel solution that builds the DSM registration task as a motion averaging problem: pair-wise DSMs are registered to build a scene graph, with edges representing relative poses between DSMs. Specifically, based on the grid structure of the large DSM, the pair-wise registration is performed using a novel nearest neighbor search method. We show that the scene graph can be optimized via an extremely fast motion average algorithm with O(N) complexity (N refers to the number of images). Evaluation of high-resolution satellite-derived DSM demonstrates significant improvement in computation and accuracy.

Figures (9)

• Figure 1: Overview of the proposed method. Given $N$ unaligned DSMs, our target is to remove their systematic errors to produce seamless registered DSMs (the unaligned and registered DSMs are color-coded based on their heights). We first construct a scene graph $\mathcal{G}(\mathcal{V},\mathcal{E})$ with edges denoting all possible pairs of DSMs.Then, the proposed DSM-ICP (see \ref{['sec:31']}) is performed for all edges determining their pair-wise transformation $\{T_{ij} | (i,j) \in \mathcal{E}\}$. Finally, a motion averaging approach is performed to estimate the global poses $\{T_i | i=1,...,N\}$ (see \ref{['sec:32']})
• Figure 2: Illustration of the proposed NN search method. To simplify, DSMs are depicted as profiles in 2D space, with the x-axis representing the horizontal plane and the y-axis denoting height. The "Blue dashed" lines depict DSM data stored on disk, while the "Green solid" line represents cached data in RAM.
• Figure 3: Comparison of two approaches for scene graph construction. "MST" denotes the minimal spanning tree, which identifies the minimum number of edges connecting all nodes, while our method establishes edges connecting all possible nodes. In this scenario, the path between any two nodes in our graph is shorter than or equal to those in the MST graph (e.g., node 4&6 in OMA3), significantly mitigating the accumulated error.
• Figure 4: Illustration of six test sets of satellite DSMs and the ground truth lidar data, where three test sets in Jacksonville area (JAX) and three test sets in Omaha area (OMA). The detailed description is in \ref{['sec:exp_dataset']}
• Figure 5: Computation efficiency of NN search for DSM-ICP and the standard ICP using k-d tree. The y-axis represents the time cost of the NN search at each iteration. In this experiment, 2065 points in total were performed NN search within the varying number of reference points from 0.5 million (a) to 305 million points (d). For ICP using k-d tree, the "Blue dashed" line represents the initialization of k-d tree, followed by NN search.