Abstract
The evolving field of mobile robotics has increased the demand for simultaneous localization and mapping (SLAM) systems. To augment localization accuracy and mapping efficacy, this research introduces a refined SLAM core module. Within the feature matching phase, cross-validation matching is employed to filter out mismatches. In the keyframe selection strategy, an exponential threshold function quantifies the selection process. Compared with single-robot systems, this multi-robot collaborative SLAM (CSLAM) system substantially improves task efficiency and robustness. Utilizing a centralized structure, the system incorporates a coarse-to-fine matching approach for multi-map point cloud registration. Built upon ORB-SLAM3, the system achieved a 12.90% reduction in absolute trajectory error across standard datasets.
Introduction
Simultaneous Localization and Mapping (SLAM) is foundational for autonomous robotics, yet state-of-the-art systems like ORB-SLAM3 face persistent challenges. Two primary hurdles include feature mismatching, where single-step verification often leads to "ghost" points and tracking errors, and heuristic keyframe selection, which typically lacks a quantitative approach to account for the complex relationships between time, movement, and view quality. Furthermore, merging maps from multiple agents often ignores prior state information, leading to high computational overhead and misaligned global maps.
System Architecture and Methodology
The proposed system enhances the standard SLAM pipeline through a centralized multi-robot architecture and refined processing modules.
Cross-Validation Feature Matching
To address the issue of outliers in visual odometry, the system implements a "reciprocity check" during feature matching. Instead of a one-way search, the system projects points from Frame B to Frame A to find matches, then performs an inverse transformation to project Frame A back to Frame B. Only points that consistently match in both directions are retained. This mathematical "handshake" significantly filters out mismatches that traditional brute-force methods might accept.
Exponential Threshold Keyframe Selection
The authors constructed a non-linear function to quantify keyframe importance, weighing four factors: temporal cost, common view quality, translation cost, and rotational cost. By utilizing an exponential function ($T_H = \sum W(e^x - 1)$), the system becomes more sensitive to rapid geometric changes, such as sharp turns. This ensures keyframes are added precisely when the motion model requires them most, preventing the "lost track" issues common in high-frequency signal scenes.
Multi-Map Registration and Processing
For multi-robot collaboration, the system uses a coarse-to-fine registration strategy. It first performs coarse matching using the robots' initial poses, followed by fine matching using the Generalized Iterative Closest Point (GICP) algorithm. To ensure the resulting dense maps are computationally viable, uniform sampling is applied (reducing data redundancy by approximately 32%) alongside 3D Gaussian filtering to smooth noise.
Experimental Results
The system was rigorously evaluated using the TUM RGB-D, EuRoC MAV, and TUM_VI datasets. Key findings include:
- Accuracy: The system outperformed ORB-SLAM3, specifically reducing the Root Mean Square Error (RMSE) by 12.90% and the maximum absolute trajectory error by 10.25% on the
fr1_desk2sequence. - Robustness: The improved keyframe strategy maintained tracking in challenging sequences like
fr3_teddy, where standard ORB-SLAM3 often struggles with sharp rotations. - Efficiency: Despite the added complexity of cross-validation, the system maintains real-time performance. Uniform sampling effectively reduced point cloud density without deteriorating mapping quality, making it suitable for real-time hardware deployment.
Conclusion and Future Work
This enhanced CSLAM system demonstrates that non-linear keyframe selection and reciprocal feature matching significantly improve localization and mapping consistency. While the current centralized architecture is robust, future research will explore distributed or hybrid schemes to mitigate the risk of single-point communication failures. Additionally, integrating deep learning to filter dynamic objects (e.g., pedestrians) remains a priority for improving performance in crowded public environments.
Summary updated