Adaptive Multi-Sensor Fusion for SLAM: A Scan Context-Driven Approach

This paper proposes a novel multi-sensor fusion SLAM algorithm, named SC-LVI-SAM, based on scanning context, to address the issues of decreased positioning accuracy caused by missing feature points and prolonged motion in complex large-scale scenes in multi-sensor fusion SLAM algorithms. Firstly, the scanning context method is used to preprocess LIDAR point cloud data, generating a descriptor of the environment. This descriptor, along with data from the IMU and vision sensors, is then used for state estimation, yielding initial pose estimates and motion information. Then, the scan context module uses the descriptor for environment recognition and loop closure detection, providing more accurate feature description and context information for fast loop closure matching. It avoids ignoring the spatial relationship and order between features due to the local feature description of DBoW2, and improves the accuracy and robustness of loop closure detection. Finally, global optimization is performed to correct accumulated errors in the entire trajectory and map. In KAIST02 and Riverside01 sequences of MulRan dataset, the root mean square error of the absolute pose error of the proposed method is reduced by 85.17% and 91.30% compared with LVI-SAM, respectively. Experimental results on multiple public benchmark datasets demonstrate that in the case of almost the same computational efficiency, the proposed algorithm effectively enhances the accuracy of positioning, the robustness of the algorithm and accuracy of mapping, improving the global consistency of the generated map.