SyS3DS: Systematic Sampling of Large-Scale LiDAR Point Clouds for Semantic Segmentation in Forestry Robotics

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Machine Learning

Background:

• 3D LiDAR sensors provide accurate environmental mapping but generate massive point clouds challenging for real-time processing.
• Existing semantic segmentation methods struggle with large datasets or rely on computationally intensive sampling, limiting practical applications in robotics.
• Efficient processing of high-resolution 3D LiDAR data is crucial for real-time robotic perception and navigation.

Purpose of the Study:

• To develop an efficient and computationally feasible method for 3D semantic segmentation of large LiDAR point clouds.
• To address the limitations of existing methods in real-time processing and feature preservation.
• To introduce SyS3DS, a systematic sampling technique that balances data reduction with geometric detail retention.

Main Methods:

• Proposed SyS3DS (Systematic Sampling for 3D Semantic Segmentation), a method based on graph coloring to select non-adjacent points.
• Incorporated local neighbor retention to preserve geometric details within the sampled subset.
• Utilized an auto-ensemble technique, passing different subsets of nodes per epoch for robust learning.
• Demonstrated processing of up to 1 million points in a single pass.

Main Results:

• SyS3DS achieves efficient semantic segmentation on large-scale datasets like Semantic3D, outperforming state-of-the-art methods.
• The method successfully preserves crucial geometric features despite significant data reduction.
• Achieved real-time processing capabilities for large 3D LiDAR point clouds.
• Preliminary study shows LiDAR-only data (intensity values) is viable for semi-autonomous robot perception.

Conclusions:

• SyS3DS offers a memory and computationally efficient solution for real-time 3D semantic segmentation of LiDAR data.
• The auto-ensemble approach enhances model robustness by leveraging diverse data subsets.
• The findings support the practical application of LiDAR-based semantic segmentation in robotics and autonomous systems.
• LiDAR intensity data alone can be sufficient for certain robot perception tasks, reducing reliance on RGB data.