MAGiC-SLAM: Multi-Agent Gaussian Globally Consistent SLAM

Vladimir Yugay,Theo Gevers,Martin R. Oswald

2024-11-25

Abstract:Simultaneous localization and mapping (SLAM) systems with novel view synthesis capabilities are widely used in computer vision, with applications in augmented reality, robotics, and autonomous driving. However, existing approaches are limited to single-agent operation. Recent work has addressed this problem using a distributed neural scene representation. Unfortunately, existing methods are slow, cannot accurately render real-world data, are restricted to two agents, and have limited tracking accuracy. In contrast, we propose a rigidly deformable 3D Gaussian-based scene representation that dramatically speeds up the system. However, improving tracking accuracy and reconstructing a globally consistent map from multiple agents remains challenging due to trajectory drift and discrepancies across agents' observations. Therefore, we propose new tracking and map-merging mechanisms and integrate loop closure in the Gaussian-based SLAM pipeline. We evaluate MAGiC-SLAM on synthetic and real-world datasets and find it more accurate and faster than the state of the art.

Computer Vision and Pattern Recognition,Artificial Intelligence,Robotics

What problem does this paper attempt to address?

This paper attempts to solve several key problems in the Multi - Agent Simultaneous Localization and Mapping (Multi - Agent SLAM) system regarding the Novel View Synthesis (NVS) ability. Specifically, these problems include:

1. **Limitations of Existing Methods**:
- Most of the existing multi - agent SLAM methods are based on distributed neural scene representations, and they have problems such as being slow, unable to accurately render real - world data, being limited to the operation of two agents, and having limited tracking accuracy.
- These methods are difficult to handle large - scale scenes and a large amount of observational data, resulting in trajectory drift and inconsistency of observations among different agents.

2. **Challenge of Global Consistency**:
- In a multi - agent environment, due to the trajectory drift and observational differences of each agent, it is very challenging to construct a globally consistent map. Existing methods are prone to cumulative errors when dealing with multiple agents, which affects the accuracy and consistency of the overall map.

3. **Lack of Flexibility**:
- Existing methods usually can only support a fixed number of agents, lacking flexibility and being unable to dynamically adjust the number of agents according to actual needs.

To solve the above problems, the paper proposes MAGiC - SLAM (Multi - Agent Gaussian Globally Consistent SLAM), a multi - agent SLAM system based on 3D Gaussian distribution. The main innovations of this system include:

- **Efficient Scene Representation**: It uses 3D Gaussian distribution to represent the scene, supports rigid - body transformation, can quickly optimize and merge maps, and significantly accelerates the SLAM process.
- **Loop - Closure Detection Mechanism**: It introduces a loop - closure detection module based on the basic visual model, which effectively improves the trajectory accuracy and can better generalize to unknown environments.
- **Flexible Tracking and Mapping Module**: Through robust tracking and loop - closure modules, it achieves superior speed and scalability and can handle any number of simultaneously operating agents.
- **Global Consistency Guarantee**: Through an optimized mapping fusion strategy, it reduces the required disk storage and processing time and ensures the consistency and accuracy of the global map.

In summary, MAGiC - SLAM aims to overcome the limitations of existing multi - agent SLAM systems in new view synthesis and provide a more efficient, accurate, and flexible solution.