Data-Efficient Decentralized Visual SLAM (1710.05772v1)

Abstract: Decentralized visual simultaneous localization and mapping (SLAM) is a powerful tool for multi-robot applications in environments where absolute positioning systems are not available. Being visual, it relies on cameras, cheap, lightweight and versatile sensors, and being decentralized, it does not rely on communication to a central ground station. In this work, we integrate state-of-the-art decentralized SLAM components into a new, complete decentralized visual SLAM system. To allow for data association and co-optimization, existing decentralized visual SLAM systems regularly exchange the full map data between all robots, incurring large data transfers at a complexity that scales quadratically with the robot count. In contrast, our method performs efficient data association in two stages: in the first stage a compact full-image descriptor is deterministically sent to only one robot. In the second stage, which is only executed if the first stage succeeded, the data required for relative pose estimation is sent, again to only one robot. Thus, data association scales linearly with the robot count and uses highly compact place representations. For optimization, a state-of-the-art decentralized pose-graph optimization method is used. It exchanges a minimum amount of data which is linear with trajectory overlap. We characterize the resulting system and identify bottlenecks in its components. The system is evaluated on publicly available data and we provide open access to the code.

• The paper proposes a data-efficient decentralized visual SLAM system integrating existing components and employing a two-stage data association process to reduce communication linearly with robot count.
• Experiments show the system significantly reduces data transmission volume compared to traditional methods while maintaining competitive accuracy parameters on public datasets.
• The integrated approach provides a viable pathway for scalable and efficient multi-robot operations in environments lacking absolute positioning by effectively managing data exchange.

Data-Efficient Decentralized Visual SLAM

The paper "Data-Efficient Decentralized Visual SLAM" by Titus Cieslewski, Siddharth Choudhary, and Davide Scaramuzza addresses the challenges posed by decentralized simultaneous localization and mapping (SLAM) for multi-robot systems in environments lacking absolute positioning. The authors have developed an innovative approach to decentralized visual SLAM, emphasizing data efficiency and scalability across robot teams.

Decentralized SLAM systems are crucial for multi-robot operations, as they alleviate the bottleneck encountered with centralized systems in computation and bandwidth. The proposed system integrates existing state-of-the-art decentralized SLAM components to create a compact yet complete decentralized visual SLAM framework. This integration brings together decentralized visual place recognition (DVPR), visual odometry (VO), relative pose estimation (RelPose), and decentralized pose graph optimization (DOpt).

The core contribution of the paper is the method's efficiency in data association and optimization. Traditional decentralized visual SLAM systems often require the exchange of large map data among robots, scaled quadratically with the number of robots. In contrast, this paper proposes a two-stage efficient data association process:

1. Stage One: A compact image descriptor, derived from NetVLAD, is deterministically sent to a single robot for place recognition, significantly reducing initial data exchanges.
2. Stage Two: Only upon successful place recognition, the data requisite for relative pose estimation is sent, again to one robot, ensuring that data transfers scale linearly with robot count.

For optimization purposes, a decentralized pose-graph optimization method is employed, minimizing data exchanges proportionally with trajectory overlap. The system effectively establishes inter-robot measurements without necessitating direct measurements or specialized hardware.

Experimental Evaluation

The results were validated through experiments on publicly available datasets like KITTI and MIT Stata Center. The paper investigates the impact of varying parameters such as the minimum distance for geometric verification and the dimensionality of NetVLAD descriptors on data transmission and trajectory accuracy.

Key findings are as follows:

• The decentralized system successfully reduces data transmission volumes while maintaining competitive accuracy parameters. For example, the adoption of visual words rather than full descriptors for feature association significantly cuts back on RelPose data transfer without degrading overall precision.
• Adjusting the minimum distance for geometric verification ($\tau_\text{mdg}$) and NetVLAD dimension offers a fine balance between communication efficiency and localization precision.
• The hierarchical optimization approach is shown to work effectively even under chain-like trajectory topologies, although convergence nuances differ when compared to grid-like structures. This opens avenues for targeted optimization based on graph topology.

Implications and Future Directions

The integration of decentralized SLAM components, as demonstrated, provides a viable pathway for scalable, efficient multi-robot operations in the absence of absolute reference frames. Looking forward, the exploration of more data-efficient association techniques and the enhancement of decentralized optimization robustness present promising areas for further research. Such advancements could enable more intricate and varied applications in unmanned systems and collaborative robotics without compromising on performance or reliability. Additionally, aligning the decentralized optimization preferences to specific graph structures can enhance efficiency and convergence in large-scale deployments.

In conclusion, the paper lays down a substantial groundwork towards achieving more data-efficient decentralized visual SLAM systems that support large heterogeneous robot teams while adapting to real-world conditions. It advances the state of decentralized SLAM, contributing towards making multi-robot systems practical and scalable in complex environments.