LCP-Fusion: A Neural Implicit SLAM with Enhanced Local Constraints and Computable Prior

Jiahui Wang, Yinan Deng, Yi Yang, Yufeng Yue

Introduction

Dense Simultaneous Localization and Mapping (SLAM) using neural implicit representations, such as those derived from Neural Radiance Fields (NeRF), has emerged as a powerful paradigm for high-fidelity 3D reconstruction. Despite their potential, current neural implicit SLAM systems face two primary bottlenecks: the requirement for pre-defined scene boundaries and the accumulation of odometry drift. These issues often manifest as "ghosting" artifacts—where surfaces appear duplicated or misaligned—and limit the utility of these systems in real-world robotic exploration of unknown environments.

To address these challenges, we present LCP-Fusion, a neural implicit SLAM system that integrates a dynamically growing sparse voxel octree with explicit Signed Distance Function (SDF) priors and a visual-overlap-based optimization strategy. By enforcing stronger local constraints, LCP-Fusion achieves superior mapping consistency and localization accuracy compared to existing state-of-the-art methods.

Methodology

1. Hybrid Scene Representation with Computable Priors

Unlike many existing systems that rely on a fixed-size neural grid, LCP-Fusion utilizes a Sparse Voxel Octree (SVO). This structure allows the system to allocate memory on-the-fly as the camera explores new territory, removing the need for prior knowledge of scene dimensions.

Each voxel vertex in the SVO stores a high-dimensional neural feature alongside a computable SDF prior. This prior is generated through a simple back-projection of depth data, providing a coarse geometric initialization. By anchoring the neural optimization to this physical reality, the system avoids the "floaty" geometry often seen in purely implicit methods, leading to faster convergence and more robust surface estimation.

2. Visual Overlap Sliding Window

Standard optimization routines in neural SLAM often sample past keyframes randomly to maintain global consistency. However, random sampling can fail to provide sufficient constraints in areas where the camera returns to a previously mapped location (loop closure).

LCP-Fusion introduces a Visual Overlap Sliding Window that selects keyframes based on actual content visibility. By re-projecting pixels from the current view into historical frames, the system identifies which past views share the most visual information. This targeted selection ensures that the local geometric "links" are strongest where they are most needed to correct drift.

3. Warping Loss and Adaptive Optimization

To further enforce consistency, LCP-Fusion employs a warping loss. This loss function ensures that a 3D point, when projected back into multiple overlapping camera views, maintains consistent color and depth values across those views.

To maintain computational efficiency, the system implements an Adaptive Early Ending policy. By monitoring the optimization loss, the system can cease iterations once a high-quality reconstruction is achieved, allowing computational resources to be redirected toward mapping newly discovered areas of the scene.

Experimental Results

LCP-Fusion was evaluated against leading baselines, including Nice-SLAM, Co-SLAM, and Vox-Fusion, across several datasets:

  • ScanNet: In real-world indoor environments characterized by noisy depth data, LCP-Fusion achieved an Average Trajectory Error (ATE) RMSE of 7.22 cm, significantly outperforming Vox-Fusion (10.04 cm) and Nice-SLAM (10.34 cm).
  • Replica: In high-quality synthetic environments, the system demonstrated superior localization and surface reconstruction accuracy.
  • Real-World Deployment: Using an Azure Kinect sensor on a mobile robot, the authors demonstrated that LCP-Fusion could successfully map unknown spaces without pre-defined bounds, effectively eliminating the "double surface" effect in loop-closure regions.

Conclusion and Future Work

LCP-Fusion demonstrates that combining sparse, scalable data structures with explicit geometric priors and targeted visual constraints can significantly mitigate the common pitfalls of neural implicit SLAM. While the system excels at geometric and photometric reconstruction, it currently lacks semantic understanding.

Future research will focus on integrating Vision-Language Models (VLMs) to enable semantic mapping, allowing robots to interpret and interact with specific objects within their environment. Additionally, extending these local constraints to work with monocular RGB input remains a key objective for making high-fidelity SLAM more accessible to low-cost sensor suites.

Summary updated

BibTeX
@inproceedings{wang2024lcp,
  author = {Jiahui Wang and Yinan Deng and Yi Yang and Yufeng Yue},
  title = {{LCP-Fusion: A Neural Implicit SLAM with Enhanced Local Constraints and Computable Prior}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2024},
  doi = {10.1109/iros58592.2024.10802626},
}
Figure