LDG-CSLAM: Multi-Robot Collaborative SLAM Based on Curve Analysis, Normal Distribution, and Factor Graph Optimization

In complex, enclosed environments where global positioning system (GPS) failures are common, multi-robot collaborative simultaneous localization and mapping (CSLAM) faces several key challenges, including redundant communication data, low fusion efficiency, and poor system robustness. These issues arise primarily due to inefficiencies in extracting and sharing descriptors of complex 3D environments, weak robustness in relative pose estimation from multiple information sources, and insufficient suppression of highly coupled dynamic estimation errors. The combined effect of these factors often leads to system failure, making it difficult to achieve stable and accurate global localization and mapping. To address these challenges, this paper proposes LDG-CSLAM, a novel multi-robot CSLAM method that integrates curve analysis, normal distribution, and factor graph optimization. LDG-CSLAM improves the efficiency of extracting and sharing global environment descriptors through key frame extraction based on point cloud curvature analysis. It further enhances performance with a distributed global mapping technique based on the normal distribution transform (NDT). Additionally, the method incorporates real-time optimization of both self and relative odometer using factor graph methods, effectively mitigating dynamic errors. This integrated design significantly reduces computational and communication overhead while improving system stability and accuracy. Experimental results, focused on operational stability, communication efficiency, and trajectory accuracy, demonstrate that LDG-CSLAM outperforms existing methods like DisCo-SLAM and DCL-SLAM, providing superior performance in multi-robot SLAM for GPS-denied environments.