A Coarse to Fine 3D LiDAR Localization with Deep Local Features for Long Term Robot Navigation in Large Environments

TL;DR

This paper tackles the problem of global localization for mobile robots with unknown initial pose in large, dynamic environments. It introduces MCL-DLF, a coarse-to-fine framework that uses MinkUNeXt to embed LiDAR point clouds into robust global descriptors for coarse localization via Monte Carlo Localization, followed by fine localization through ICP or deep local features extracted from MinkUNeXt intermediate layers. The main contributions are a complete MCL-DLF pipeline, a comparative analysis against ICP and state-of-the-art methods, and extensive validation on the NCLT and UMH datasets demonstrating robustness to seasonal changes and indoor–outdoor handovers. The approach supports accurate, long-term navigation in challenging environments and provides publicly available code for reproducibility.

Abstract

The location of a robot is a key aspect in the field of mobile robotics. This problem is particularly complex when the initial pose of the robot is unknown. In order to find a solution, it is necessary to perform a global localization. In this paper, we propose a method that addresses this problem using a coarse-to-fine solution. The coarse localization relies on a probabilistic approach of the Monte Carlo Localization (MCL) method, with the contribution of a robust deep learning model, the MinkUNeXt neural network, to produce a robust description of point clouds of a 3D LiDAR within the observation model. For fine localization, global point cloud registration has been implemented. MinkUNeXt aids this by exploiting the outputs of its intermediate layers to produce deep local features for each point in a scan. These features facilitate precise alignment between the current sensor observation and one of the point clouds on the map. The proposed MCL method incorporating Deep Local Features for fine localization is termed MCL-DLF. Alternatively, a classical ICP method has been implemented for this precise localization aiming at comparison purposes. This method is termed MCL-ICP. In order to validate the performance of MCL-DLF method, it has been tested on publicly available datasets such as the NCLT dataset, which provides seasonal large-scale environments. Additionally, tests have been also performed with own data (UMH) that also includes seasonal variations on large indoor/outdoor scenarios. The results, which were compared with established state-of-the-art methodologies, demonstrate that the MCL-DLF method obtains an accurate estimate of the robot localization in dynamic environments despite changes in environmental conditions. For reproducibility purposes, the code is publicly available at https://github.com/miriammaximo/MCL-DLF.git

Figures (8)

• Figure 1: Schematic of the proposed localization method, consisting of coarse localization and fine localization. Coarse localization process: The inputs are the point clouds of the map and the query point cloud, which is derived from the current observation of the robot. From these points clouds the global descriptors are extracted with MinkUNeXt. These descriptors are used as an observation model in the MCL process. Finally, the output of this process is the estimated (x,y) position of the robot. Fine localization process: Two alternative methods are presented, both requiring as input the query point cloud and the closest point cloud of the map to the position (x,y) obtained with MCL. The first alternative method, is ICP, which also requires an initial transformation to relate the two point clouds in position and orientation. These data are obtained from the coarse localization process. The second alternative method, involves the use of an intermediate layer (3D Sparse Transpose Convolution 2) of the MinkUNeXt network to obtain the features of each point of the scans F and Q, which are used to do feature matching between points in both scans. The outputs of these two methods are the 6DoF estimated position of the robot.
• Figure 2: Weighting process of particle i. (a) the map nodes, located at positions $\{(n_{1,x}, n_{1,y}),...,(n_{N,x}, n_{N,y})\}$, are visualized as grey circles. Particle i, at coordinates $(x_i, y_i)$, is shown in red, and the robot, for which we only know the descriptor $d_{query}$ from the LiDAR observation, is shown in green. This descriptor is used to compare with all map descriptors $\{d_1,..,d_N\}$. (b) the map node with the closest descriptor is shown as a grey circle. Then, $v_j$ is calculated as the difference between the node's position $(n_{j,x}, n_{j,y})$ and the particle's position $(x_i, y_i)$. Finally, $h_j$ is calculated as the difference between $d_k$ and $d_i$, where $d_i$ is the descriptor of the map node where the particle $i$ is located.
• Figure 3: Particle locations, in red, relative to the map positions, in grey, and the robot's current location, in green. (a) the particles are shown distributed across all map nodes. (b) and (c) the particles converge to a smaller area of the map. (d) the particles have converged to very close areas near the robot's actual location.
• Figure 4: Appearance of the dataset environments captured at UMH. (a) (b) (c) outdoor environments and (d) (e) (f) indoor environments. (c) a trajectory captured entirely in an outdoor environment is shown. (f) a trajectory captured partly outdoors and partly indoors is presented.
• Figure 5: Aggregated point clouds of a robot trajectory in the campus of Miguel Hernandez University of Elche.