Abstract:Simultaneously odometry and mapping using LiDAR data is an important task for mobile systems to achieve full autonomy in large-scale environments. However, most existing LiDAR-based methods prioritize tracking quality over reconstruction quality. Although the recently developed neural radiance fields (NeRF) have shown promising advances in implicit reconstruction for indoor environments, the problem of simultaneous odometry and mapping for large-scale scenarios using incremental LiDAR data remains unexplored. To bridge this gap, in this paper, we propose a novel NeRF-based LiDAR odometry and mapping approach, NeRF-LOAM, consisting of three modules neural odometry, neural mapping, and mesh reconstruction. All these modules utilize our proposed neural signed distance function, which separates LiDAR points into ground and non-ground points to reduce Z-axis drift, optimizes odometry and voxel embeddings concurrently, and in the end generates dense smooth mesh maps of the environment. Moreover, this joint optimization allows our NeRF-LOAM to be pre-trained free and exhibit strong generalization abilities when applied to different environments. Extensive evaluations on three publicly available datasets demonstrate that our approach achieves state-of-the-art odometry and mapping performance, as well as a strong generalization in large-scale environments utilizing LiDAR data. Furthermore, we perform multiple ablation studies to validate the effectiveness of our network design. The implementation of our approach will be made available at <a class="link-external link-https" href="https://github.com/JunyuanDeng/NeRF-LOAM" rel="external noopener nofollow">this https URL</a>.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: in large - scale outdoor environments, how to use LiDAR (Light Detection and Ranging) data to simultaneously achieve high - precision odometry and high - quality mapping. Most of the existing LiDAR - based methods focus more on tracking quality and neglect reconstruction quality. Especially in large - scale scenarios, the research on simultaneously performing high - precision odometry and high - quality mapping is still insufficient. Specifically, the paper points out the following problems: 1. **Limitations of existing methods**: - Most of the existing LiDAR odometry and mapping algorithms focus on improving tracking quality while ignoring the accurate capture and high - fidelity reconstruction of the environmental geometric structure. - Although Neural Radiance Field (NeRF) has shown the potential of implicit reconstruction in indoor environments, its application to the Simultaneous Localization and Mapping (SLAM) task of large - scale incremental LiDAR data has not been fully explored. 2. **Challenges in large - scale outdoor environments**: - The sparsity of LiDAR data and the lack of RGB information make it difficult to apply the existing deep - learning algorithms to large - scale outdoor environments. - Traditional LiDAR odometry and mapping algorithms (such as the ICP algorithm) can provide accurate odometry estimates, but the generated maps are relatively rough. 3. **Need for new solutions**: - A new method that can simultaneously achieve high - precision odometry and high - quality mapping in large - scale outdoor environments is needed to meet the requirements of applications such as autonomous driving systems. To solve these problems, the paper proposes a NeRF - based LiDAR odometry and mapping method - NeRF - LOAM. This method achieves high - precision odometry and high - quality mapping in large - scale outdoor environments by introducing neural implicit representation, dynamic voxel embedding generation strategy and key - scan optimization strategy. ### Main contributions 1. **First proposed**: As far as the authors know, NeRF - LOAM is the first neural implicit odometry and mapping method for using LiDAR data in large - scale outdoor environments. 2. **Novel neural SDF module**: A neural signed distance function (SDF) module that combines dynamic generation and key - scan optimization strategies is proposed, achieving rapid voxel embedding assignment and high - fidelity 3D representation. 3. **No pre - training required**: Based on online joint optimization, this method does not require pre - training and has good generalization ability in different environments. ### Method overview NeRF - LOAM consists of three main modules: neural odometry, neural mapping and mesh reconstruction. These modules jointly use the proposed neural signed distance function (SDF) to separate LiDAR point clouds into ground points and non - ground points, thereby reducing Z - axis drift and generating smooth and dense 3D maps. In addition, this method also designs a dynamic voxel embedding generation strategy and a key - scan optimization strategy to improve the reconstruction quality and the ability to handle incremental data. ### Experimental results The experimental results show that NeRF - LOAM achieves state - of - the - art odometry and mapping performance on multiple public datasets and shows strong generalization ability in large - scale outdoor environments.

NeRF-LOAM: Neural Implicit Representation for Large-Scale Incremental LiDAR Odometry and Mapping

Neural Implicit Representation for Highly Dynamic LiDAR Mapping and Odometry

RaLL: End-to-end Radar Localization on Lidar Map Using Differentiable Measurement Model

VI-NeRF-SLAM: a real-time visual–inertial SLAM with NeRF mapping

LISNeRF Mapping: LiDAR-based Implicit Mapping via Semantic Neural Fields for Large-Scale 3D Scenes

Neural Surface Reconstruction and Rendering for LiDAR-Visual Systems

Efficient LiDAR Odometry for Autonomous Driving.

LiDAR-NeRF: Novel LiDAR View Synthesis via Neural Radiance Fields

LiDeNeRF: Neural radiance field reconstruction with depth prior provided by LiDAR point cloud

NeSLAM: Neural Implicit Mapping and Self-Supervised Feature Tracking With Depth Completion and Denoising

NF-Atlas: Multi-Volume Neural Feature Fields for Large Scale LiDAR Mapping

NeRF-LiDAR: Generating Realistic LiDAR Point Clouds with Neural Radiance Fields

SAR-SLAM: Self-Attentive Rendering-based SLAM with Neural Point Cloud Encoding

Camera and LiDAR Fusion for Urban Scene Reconstruction and Novel View Synthesis via Voxel-Based Neural Radiance Fields

NeRF-VO: Real-Time Sparse Visual Odometry with Neural Radiance Fields

RF-LOAM: Robust and Fast LiDAR Odometry and Mapping in Urban Dynamic Environment

CLONeR: Camera-Lidar Fusion for Occupancy Grid-aided Neural Representations

SLAM Meets NeRF: A Survey of Implicit SLAM Methods

Loc-NeRF: Monte Carlo Localization using Neural Radiance Fields

Ternary-Type Opacity and Hybrid Odometry for RGB NeRF-SLAM

PC-NeRF: Parent-Child Neural Radiance Fields Using Sparse LiDAR Frames in Autonomous Driving Environments