Density-equalized RANDT scan matching with integrated outlier removal and point density uniformity

Area of Science:

• Computer Vision
• Robotics
• Geospatial Analysis

Background:

• 3D point cloud data is crucial for environmental sensing, geospatial analysis, and autonomous driving.
• Accurate registration of point cloud data is essential for noise reduction and uniform point density.
• Existing registration algorithms struggle with slow convergence, partial overlaps, and local minima due to non-homogeneous feature points.

Purpose of the Study:

• To introduce a novel scan matching framework, Robust and Adaptive Normal Distribution Transform (RANDT), to address limitations in 3D point cloud registration.
• To improve the accuracy, density, and uniformity of point cloud maps.

Main Methods:

• Implemented an incremental scan matching module for continuous matching with new scans.
• Integrated an outlier removal feature to eliminate noisy data points before scan matching.
• Developed the Robust and Adaptive Normal Distribution Transform (RANDT) framework.

Main Results:

• RANDT achieved Root Mean Square Error (RMSE) values of 0.054 m and 0.062 m on KITTY and ModelNet40 datasets.
• Demonstrated an 18-25% error reduction, even with noisy data and partial overlaps.
• Attained the highest point density uniformity score of 0.92 compared to baseline methods.

Conclusions:

• The proposed RANDT method offers a robust and adaptive solution for 3D point cloud registration.
• RANDT effectively handles noise and partial overlaps, leading to more accurate and uniform point cloud maps.
• This advancement is significant for applications like autonomous driving and environmental sensing.