Online Mapping from Weight Matching Odometry and Highly Dynamic Point Cloud Filtering via Pseudo-Occupancy Grid

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Geospatial Data Processing

Background:

• Autonomous systems require precise mapping and localization for efficient operation.
• Existing methods struggle with accurately identifying and filtering dynamic objects in real-time.

Purpose of the Study:

• To develop a high-accuracy online mapping system integrating LiDAR-IMU-GNSS odometry.
• To implement an object-level dynamic point cloud filtering method for improved map clarity.

Main Methods:

• Utilized IMU pre-integration, ground point segmentation (PMF), motion compensation, and weight feature point matching for odometry.
• Employed a pseudo-occupancy grid to identify dynamic grids and curved voxel clustering for object segmentation.
• Integrated these components for real-time, high-accuracy online mapping with dynamic object filtering.

Main Results:

• The proposed odometry achieved superior accuracy on KITTI, UrbanLoco, and NCD datasets compared to LIO-SAM and FAST-LIO2.
• The dynamic point cloud filtering algorithm demonstrated higher detection precision than Removert and ERASOR.
• Successfully generated a real-time map with comprehensive filtering of vehicles, cyclists, and pedestrians.

Conclusions:

• The developed system offers a significant advancement in high-accuracy online mapping for autonomous applications.
• The novel filtering method effectively removes highly dynamic objects, enhancing map reliability.
• This research contributes to safer and more efficient autonomous navigation through improved perception and mapping.