NDT-Map-Code: A 3D global descriptor for real-time loop closure detection in lidar SLAM

TL;DR

This work tackles loop-closure in lidar SLAM under sparse or limited-FOV conditions by introducing NDT-MC, a global descriptor built from lightweight NDT maps. It combines multi-layer Polar-Range-Height ROI partitioning with explicit NDT shape classification and entropy, encoding into a compact descriptor and a fast GK-based matching pipeline that remains rotation-invariant via sector keys. Evaluations on NIO underground parking data and KITTI show superior accuracy and real-time performance, including integration with LIO-SAM for full SLAM in real time. The method supports crowd-sourced mapping with low maintenance and computational overhead, and the authors provide public code for reproducibility and community use.

Abstract

Loop-closure detection, also known as place recognition, aiming to identify previously visited locations, is an essential component of a SLAM system. Existing research on lidar-based loop closure heavily relies on dense point cloud and 360 FOV lidars. This paper proposes an out-of-the-box NDT (Normal Distribution Transform) based global descriptor, NDT-Map-Code, designed for both on-road driving and underground valet parking scenarios. NDT-Map-Code can be directly extracted from the NDT map without the need for a dense point cloud, resulting in excellent scalability and low maintenance cost. The NDT representation is leveraged to identify representative patterns, which are further encoded according to their spatial location (bearing, range, and height). Experimental results on the NIO underground parking lot dataset and the KITTI dataset demonstrate that our method achieves significantly better performance compared to the state-of-the-art.

Figures (6)

• Figure 1: Polar-range-height coordinates based ROI partitioning. We first divide the 3D space according to ring, sector, and height, and the corresponding polar-range-height coordinate axes, i.e. $r$, $\theta$, and $w$, can be obtained. Afterward, the ring sector corresponding to different heights ($w$) is transformed into a $Cartesian$ coordinate system with $\theta$ as the abscissa and $r$ as the ordinate.
• Figure 2: The overview of NDT-MC. The construction of NDT-MC involves the following steps: 1. Converting point clouds into NDT representation. 2. Calculating geometric values and entropy for each NDT cell. 3. Computing the coordinates of each NDT cell in the Polar-Range-Height coordinate system based on its mean value. 4. Constructing the proposed descriptor using a strengthened height-layer encoding with geometric and entropy components.
• Figure 3: The geometrical shape of NDT cells corresponding to different $g$ values. Four subfigures illustrate representative shapes for corresponding $g$ values, namely plane, ellipsoid, sphere, and line. In each subfigure, a 3D NDT cell is shown along with a 2D projection in the X, Y, and Z directions.
• Figure 4: Construction of geometric key. $s$ is a segmented parameter for the $g$ value.
• Figure 5: A visualization of trajectories of the experimental dataset. Each sub-figure has three trajectories of different colors, in which blue, red, and black represent the trajectories of test data 1, test data 2, and the database respectively. Trajectories of the same underground parking-lot were collected at different times over different days.