Improving Indoor Localization Accuracy by Using an Efficient Implicit Neural Map Representation

TL;DR

This work tackles the challenge of accurate and efficient global localization for indoor robots by replacing traditional occupancy grids with an implicit neural map (ENM) that learns surface geometry from 2D LiDAR. ENM combines a dense learnable feature grid with a lightweight MLP to predict both non-projective and direction-aware signed distance fields, which are then used in a standard Monte Carlo Localization framework to update particle weights. The authors demonstrate state-of-the-art localization accuracy, real-time pose tracking, and robustness to large particle counts on indoor sequences, along with favorable runtime characteristics. The contributions include the ENM representation, its integration into ENM-MCL, and a thorough experimental evaluation showing significant performance gains and practical feasibility for real-time robotic localization in complex indoor environments.

Abstract

Globally localizing a mobile robot in a known map is often a foundation for enabling robots to navigate and operate autonomously. In indoor environments, traditional Monte Carlo localization based on occupancy grid maps is considered the gold standard, but its accuracy is limited by the representation capabilities of the occupancy grid map. In this paper, we address the problem of building an effective map representation that allows to accurately perform probabilistic global localization. To this end, we propose an implicit neural map representation that is able to capture positional and directional geometric features from 2D LiDAR scans to efficiently represent the environment and learn a neural network that is able to predict both, the non-projective signed distance and a direction-aware projective distance for an arbitrary point in the mapped environment. This combination of neural map representation with a light-weight neural network allows us to design an efficient observation model within a conventional Monte Carlo localization framework for pose estimation of a robot in real time. We evaluated our approach to indoor localization on a publicly available dataset for global localization and the experimental results indicate that our approach is able to more accurately localize a mobile robot than other localization approaches employing occupancy or existing neural map representations. In contrast to other approaches employing an implicit neural map representation for 2D LiDAR localization, our approach allows to perform real-time pose tracking after convergence and near real-time global localization. The code of our approach is available at: https://github.com/PRBonn/enm-mcl.

Figures (5)

• Figure 1: We learn a neural network to represent the surfaces of environments. Our method achieves both efficiency and accuracy for indoor localization by integrating our proposed efficient neural map representation into a Monte Carlo localization system. We show the average runtime for each method in sequence 1 of the in-house dataset.
• Figure 2: Overview of our approach for jointly predicting the non-projective SDF and direction-aware projective SDF values using our proposed efficient neural map representation. We sample several positions along a LiDAR ray and input the 2D position ${\hbox{\boldmath$p$}} = (x, y)^{\top}$ as well as the corresponding ray direction ${\hbox{\boldmath$d$}} = (d_{x}, d_{y})^{\top}$ into our ENM model to estimate the SDF and PSDF. The predictions of the SDF and PSDF values by the neural network are supervised with the ground truth SDF/PSDF values from 2D LiDAR measurements.
• Figure 3: The observation model based on the ENM representation. We estimate the SDF and PSDF values for all beam end-points of the particles using the ENM model, and update the particles' weights by computing the likelihood based on the SDF and PSDF values of the beam end-points. The positive particle has higher likelihood which scans are aligned with the zero level of distance field.
• Figure 4: Qualitative global localization results of ENM-MCL on all five sequences of the in-house dataset. We show the predicted trajectories after convergence. The predicted trajectories are mostly aligned with the ground truth, indicating the high accuracy and reliability of our method.
• Figure 5: The error curves of location RMSE on the Sequence 5 of the in-house dataset.