CELLmap: Enhancing LiDAR SLAM through Elastic and Lightweight Spherical Map Representation

TL;DR

This work designs an elastic and lightweight map representation called CELLmap, composed of several CELLS, each representing the local map at the corresponding location, and designs a general backend including CELL-based bidirectional registration module and loop closure detection module to improve global map consistency.

Abstract

SLAM is a fundamental capability of unmanned systems, with LiDAR-based SLAM gaining widespread adoption due to its high precision. Current SLAM systems can achieve centimeter-level accuracy within a short period. However, there are still several challenges when dealing with largescale mapping tasks including significant storage requirements and difficulty of reusing the constructed maps. To address this, we first design an elastic and lightweight map representation called CELLmap, composed of several CELLs, each representing the local map at the corresponding location. Then, we design a general backend including CELL-based bidirectional registration module and loop closure detection module to improve global map consistency. Our experiments have demonstrated that CELLmap can represent the precise geometric structure of large-scale maps of KITTI dataset using only about 60 MB. Additionally, our general backend achieves up to a 26.88% improvement over various LiDAR odometry methods.

Figures (5)

• Figure 1: Performance of CELLmap on KITTI 00 sequence. (a) The CELLmap of KITTI 00 consists of 577 CELLs. (b) The corresponding point cloud map. (c) and (d) are zoomed-in views of (a) and (b), respectively. (e) The CELL is extremely lightweight, requiring only about 100KB of storage per CELL, with the entire KITTI 00 scene needing just 60.2MB. (f) and (g) illustrate a general backend to optimize the CELLmap. (f) By introducing more constraints through CELL-based bidirectional registration, drift over short distances is reduced, thereby improving pose estimation accuracy. (g) Enhancing global map consistency through loop closure detection.
• Figure 2: The overview for constructing CELLmap. We first build a local point cloud map using the output from any LiDAR odometry. Then, we generate the corresponding CELL through the CELL generation module. Next, we enhance the trajectory accuracy of the utilized LiDAR odometry through CELL-based bidirectional registration and loop detection algorithms. Finally, we output a globally consistent CELLmap.
• Figure 3: CELL generation. We first partition the unit sphere using a unique spherical point cloud $\mathcal{FL}$. Then, each time a CELL needs to be generated, the corresponding local point cloud map is segmented according to $\mathcal{FL}$, and the plane parameters for each segment are extracted. Finally, each CELL is composed of an $N_{sp}*4$ array.
• Figure 4: An example of CELL generation. Each local point cloud map is divided into segments according to $\mathcal{FL}$ shown in (b). The segmentation result is shown in (c), with different colors representing different segments. Then, plane extraction is performed for each segment to calculate its normal. We display the segments with valid planes in (d), colored according to their normal vectors.
• Figure 5: Scan-to-CELL registration. First, the planes in the map are reconstructed according to the plane attributes and spherical point cloud $\mathcal{FL}$. Correspondences between LiDAR points and planes are established by finding the nearest points in $\mathcal{FL}$. Then, point-to-plane registration is performed to solve for the LiDAR frame's pose.