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

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Existing LiDAR SLAM methods often treat point cloud processing and pose optimization as separate stages.
• Fixed settings in traditional SLAM limit performance with changing environments or motion.
• This can lead to suboptimal computational efficiency and localization accuracy.

Purpose of the Study:

• To develop an adaptive LiDAR-inertial SLAM framework with bidirectional closed-loop coupling.
• To enhance localization accuracy and mapping quality by integrating adaptive point cloud processing with pose optimization.
• To address the limitations of loosely connected stages in existing SLAM methods.

Main Methods:

• Implemented a LiDAR-inertial SLAM framework with adaptive frontend point cloud processing and backend pose optimization.
• Dynamically adjusted depth image resolution based on pose uncertainty and loop closure importance.
• Utilized a comprehensive score for selecting high-quality sparse points and incorporated it into uncertainty-weighted ICP and factor graph optimization.

Main Results:

• Achieved a 15.8% and 15.2% reduction in mean RMSE on KITTI and M2DGR datasets compared to FAST-LIO2.
• Demonstrated a 26.3% RMSE reduction in real-world field tests.
• Showcased significant improvements in both mapping quality and localization accuracy.

Conclusions:

• The proposed adaptive LiDAR-inertial SLAM framework effectively improves localization accuracy and mapping quality.
• Bidirectional closed-loop coupling enhances robustness in dynamic environments.
• The adaptive approach optimizes computation and performance for robotic navigation and autonomous driving.