Abstract
To address the challenges in multi-robot collaborative SLAM, including excessive redundant computations and low processing efficiency in candidate loop closure selection during front-end loop detection, as well as high computational complexity and long iteration times due to global pose optimization in the back-end, this paper introduces several key improvements. First, a global matching and candidate loop selection strategy is incorporated into the front-end loop detection module, leveraging both LiDAR point clouds and visual features to achieve cross-robot loop detection, effectively mitigating computational redundancy and reducing false matches in collaborative multi-robot systems. Second, an improved distributed robust pose graph optimization algorithm is proposed in the back-end module. By introducing a robust cost function to filter out erroneous loop closures and employing a subgraph optimization strategy during iterative optimization, the proposed approach enhances convergence speed and solution quality, thereby reducing uncertainty in multi-robot pose association. Experimental results demonstrate that the proposed method significantly improves computational efficiency and localization accuracy. Specifically, in front-end loop detection, the proposed algorithm achieves an F1-score improvement of approximately 8.5-51.5% compared to other methods. In back-end optimization, it outperforms traditional algorithms in terms of both convergence speed and optimization accuracy. In terms of localization accuracy, the proposed method achieves an improvement of approximately 32.8% over other open source algorithms.