LIO-CSI: LiDAR inertial odometry with loop closure combined with semantic information.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• State estimation and mapping are crucial for autonomous driving.
• Geometric feature matching for pose estimation is vulnerable to dynamic objects.
• Deep learning enables semantic information extraction from point clouds.

Purpose of the Study:

• To develop a robust LiDAR inertial odometry system that integrates semantic information.
• To improve the accuracy of point cloud registration and loop closure detection in dynamic environments.
• To enhance the reliability of autonomous driving systems.

Main Methods:

• Optimized semantic labels from Sparse Point-Voxel Neural Architecture Search (SPVNAS).
• Integrated semantic information into tightly-coupled LiDAR inertial odometry via smoothing and mapping (LIO-SAM) to filter dynamic objects.
• Developed a semantic-assisted scan-context method for loop closure detection.

Main Results:

• The proposed LIO-CSI method significantly improves accuracy and robustness compared to purely geometric methods.
• Effective filtering of dynamic objects like pedestrians and vehicles enhances point cloud registration.
• Improved performance in loop closure detection, particularly in scenarios with dynamic elements.

Conclusions:

• Integrating semantic information into LiDAR inertial odometry is effective for enhancing autonomous driving systems.
• The LIO-CSI method demonstrates strong generalization capabilities across different datasets.
• Semantic-assisted approaches offer a promising direction for robust state estimation in real-world autonomous driving scenarios.