Range-FDSeg: LiDAR semantic segmentation based on fusion interactive learning and dynamic sampling for autonomous...

Area of Science:

• Computer Vision
• Robotics
• Geospatial Data Analysis

Background:

• LiDAR semantic segmentation is crucial for autonomous driving and robot navigation, improving scene perception and object detection.
• Existing methods struggle to balance high segmentation accuracy with low computational cost and complexity.
• Projecting 3D point clouds to 2D range images risks information loss and compression.

Purpose of the Study:

• To propose an efficient and accurate LiDAR semantic segmentation network, Range-FDSeg.
• To address the challenges of information loss and noise in 3D to 2D projection.
• To develop a lightweight and dynamic upsampling method for improved performance.

Main Methods:

• Introduced a multi-channel fusion interactive learning (FIL) module integrating coordinates, depth, and reflectivity.
• Designed a lightweight and dynamic upsampler, Dysample-S+, with an adaptive weighting mechanism.
• Evaluated the network on SemanticKITTI, SemanticPOSS, and NuScenes benchmark datasets.

Main Results:

• The FIL module effectively reduces noise and captures inter-channel relationships.
• Dysample-S+ dynamically adjusts to point cloud variations, overcoming traditional sampling limitations.
• Range-FDSeg demonstrated superior performance compared to state-of-the-art methods on benchmark datasets.

Conclusions:

• Range-FDSeg offers an efficient and accurate solution for LiDAR semantic segmentation.
• The proposed FIL and Dysample-S+ modules significantly enhance segmentation performance.
• The network shows strong potential for real-world applications in autonomous systems.