Cheesemap: A High-Performance Point-Indexing Data Structure for Neighbor Search in LiDAR Data

TL;DR

The paper tackles efficient neighbor search in large LiDAR point clouds by benchmarking multiple data structures and introducing cheesemap, a grid-based voxel indexing approach. Cheesemap enables fast fixed-radius and $k$-NN queries via dense, sparse, and mixed variants that adapt to density, while maintaining a competitive memory footprint. Results show cheesemap often outperforms state-of-the-art structures on ALS datasets and remains competitive across other datasets, with performance depending on density and dimensionality. The work provides practical guidance for selecting point-cloud indexing structures and outlines concrete avenues for further optimization.

Abstract

Point cloud data, as the representation of three-dimensional spatial information, is a fundamental piece of information in various domains where indexing and querying these point clouds efficiently is crucial for tasks such as object recognition, autonomous navigation, and environmental modeling. In this paper, we present a comprehensive comparative analysis of various data structures combined with neighboring search methods across different types of point clouds. Additionally, we introduce a novel data structure, cheesemap, to handle 3D LiDAR point clouds. Exploring the sparsity and irregularity in the distribution of points, there are three flavors of the cheesemap: dense, sparse, and mixed. Results show that the cheesemap can outperform state-of-the-art data structures in terms of execution time per query, particularly for ALS (Aerial Laser Scanning) point clouds. Memory consumption is also minimal, especially in the sparse and mixed representations, making the cheesemap a suitable choice for applications involving three-dimensional point clouds.

Figures (12)

• Figure 1: Spherical query (2.5 radius) in a synthetic point cloud of dimensions 100 x 100 x 50 on a system with a 6-core (12-threads) Intel Core i7-9750H and 32 DDR4 memory at 2666.
• Figure 2: Example of sparsity in a point cloud. There are no points within the light-gray ellipse (representing, for example, a lake in ALS point clouds), so shaded voxels do not contain any points.
• Figure 3: Example of kernel-based search. The red circle represents the kernel, a sphere centered in $c$ with radius $r$. The red dashed box represents the bounding box of the sphere. In light gray, the voxels that intersect with the bounding box, and may intersect with the kernel. In dark gray, those voxels that actually intersect with the sphere.
• Figure 4: Geometric average of the normalized search time of the cheesemap against nanoflann::KdTree in the spherical-kernel search. Lower is better.
• Figure 5: Geometric average of the normalized search time of the cheesemap against nanoflann::KdTree in the cube-kernel search. Lower is better.