Feedback-Driven SLAM with Adaptive Point Cloud Selection and Uncertainty-Aware Pose Optimization

LiDAR SLAM is widely used in robotic navigation and autonomous driving, but many existing methods still handle frontend point cloud processing and backend pose optimization as two loosely connected stages with fixed settings. This can lead to unnecessary computation and also limits the localization performance when the environment or motion changes. To address this issue, we propose a LiDAR–inertial SLAM framework with bidirectional closed-loop coupling between adaptive point cloud processing and pose optimization. In the frontend, depth image resolution is adjusted online according to backend pose uncertainty and loop closure importance, and a comprehensive score integrating point density, depth stability, geometric complexity, and motion consistency is used to select high-quality sparse points. In the backend, the comprehensive score is further combined with depth image quantization error to construct per-point covariance matrices for uncertainty-weighted scan-to-map ICP and factor graph noise modeling. Experiments on the KITTI and M2DGR datasets show that the proposed method reduced the mean RMSE by 15.8% and 15.2%, respectively, compared with FAST-LIO2, while the real-world field test further shows a 26.3% RMSE reduction with respect to the constructed reference trajectory. These results show that the proposed framework improves both mapping quality and localization accuracy.