L-PR: Exploiting LiDAR Fiducial Marker for Unordered Low Overlap Multiview Point Cloud Registration

TL;DR

L-PR introduces a robust framework for unordered low-overlap multiview point cloud registration by exploiting LiDAR fiducial markers. It combines an adaptive threshold marker detection pipeline with a two-level MAP optimization: a first-level weighted graph provides efficient initial pose values via shortest-path propagation, followed by a second-level factor graph that jointly refines scan poses, marker poses, and marker corners using Gauss-Newton-style optimization. The approach achieves superior registration accuracy and efficiency on challenging scenes, enables practical applications such as 3D asset collection, training data augmentation, degraded-scene reconstruction, GPS-denied localization, and large-scale map merging, and extends the training corpus with the Livox-3DMatch dataset. The work highlights the practicality of thin-sheet LiDAR fiducials for scalable, low-cost 3D data acquisition and learning-based method enhancement, supported by extensive experiments and ablation studies.

Abstract

Point cloud registration is a prerequisite for many applications in computer vision and robotics. Most existing methods focus on pairwise registration of two point clouds with high overlap. Although there have been some methods for low overlap cases, they struggle in degraded scenarios. This paper introduces a novel framework dubbed L-PR, designed to register unordered low overlap multiview point clouds leveraging LiDAR fiducial markers. We refer to them as LiDAR fiducial markers, but they are the same as the popular AprilTag and ArUco markers, thin sheets of paper that do not affect the 3D geometry of the environment. We first propose an improved adaptive threshold marker detection method to provide robust detection results when the viewpoints among point clouds change dramatically. Then, we formulate the unordered multiview point cloud registration problem as a maximum a-posteriori (MAP) problem and develop a framework consisting of two levels of graphs to address it. The first-level graph, constructed as a weighted graph, is designed to efficiently and optimally infer initial values of scan poses from the unordered set. The second-level graph is constructed as a factor graph. By globally optimizing the variables on the graph, including scan poses, marker poses, and marker corner positions, we tackle the MAP problem. We conduct both qualitative and quantitative experiments to demonstrate that the proposed method surpasses previous state-of-the-art (SOTA) methods and to showcase that L-PR can serve as a low-cost and efficient tool for 3D asset collection and training data collection. In particular, we collect a new dataset named Livox-3DMatch using L-PR and incorporate it into the training of the SOTA learning-based method, SGHR, which brings evident improvements for SGHR on various benchmarks.

Figures (15)

• Figure 1: (a): Overview of L-PR. Given ① an unordered set of unaligned, low overlap 3D point clouds, we aim to register them into ⑤ a complete point cloud utilizing LiDAR fiducial markers (thin sheets of paper/board attached to other planes). This is achieved through the proposed ② adaptive threshold marker detection method and the ③ ④ two-level graphs. The ③ first-level graph handles the unordered set of point clouds and calculates initial values. The ④ second-level graph registers the point clouds by finding the optimal solution to a maximum a-posteriori problem. (b): The proposed method can serve as a convenient, efficient, and low-cost tool for diverse applications, including 3D asset collection from sparse observations, collecting training data in novel scenes, reconstruction of degraded scenes, localization in GPS-denied environments, and merging large-scale low overlap 3D maps.
• Figure 2: (a) Typical calibration board: forms a thick board with holes and/or regions with high-intensity materials. It impacts the 3D environment and does not support patterns generated by an encoding-decoding algorithm. (b) AprilTag ap3: has rich patterns generated from the encoding-decoding algorithm and is thus robust to false positives/negatives. It has no effect on the 3D environment but is designed for cameras. (c) LiDARTag lt: has patterns compatible with AprilTag ap3 but still forms an additional 3D object. (d) IFM iilfm: has patterns compatible with AprilTag ap3 and ArUco aruco. It has no impact on the 3D environment.
• Figure 3: The raw intensity image binarized with different threshold values.
• Figure 4: An illustration of the first-level graph. After applying the proposed adaptive marker detection to all scans, an exhaustive weighted graph is constructed, with the scans and markers as nodes and the point-to-point errors as edge weights. The aim is to derive the relative pose between each non-anchor scan and the anchor scan with optimal accuracy. However, for a given non-anchor scan, such as $f_{2}$ in this simple case, there may be multiple possible paths in the exhaustive graph leading to the anchor scan ($f_{1}$). Thus, Dijkstra’s algorithm dij is employed to find the optimal path with the minimum accumulation of point-to-point errors (weights).
• Figure 5: The procedures for formulating the second-level graph. The variable nodes are represented by circles and the factor nodes are represented by squares. In Stage One, when a marker is detected in a scan, six types of nodes are added to the graph, including (1) scan pose in $\{G\}$, (2) marker pose in $\{G\}$, (3) corner positions in $\{G\}$, (4) marker pose from $\{M\}$ to $\{G\}$, (5) corner positions in $\{F\}$, and (6) corner positions in $\{M\}$, along with their corresponding edges. In Stage Two, all the marker detection results are traversed, and the operation from Stage One is repeated for each detected marker. In Stage Three, a prior factor connecting the anchor scan is added, along with factors representing the relative poses between the anchor scan and non-anchor scans.