HyperG-PS: Voxel correlation modeling via hypergraph for LiDAR panoptic segmentation.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• LiDAR point cloud panoptic segmentation is crucial for autonomous driving, integrating static and dynamic object recognition.
• Current methods face challenges in accurately clustering instances within complex LiDAR data.

Purpose of the Study:

• To propose a novel bottom-up panoptic segmentation framework for LiDAR point clouds.
• To enhance the performance of instance segmentation clustering using hypergraph learning.

Main Methods:

• Developed HyperG-PS, a framework utilizing multi-view feature extraction to fuse 3D point cloud and 2D Bird's Eye View (BEV) features.
• Modeled voxel correlations with a hypergraph to bridge voxel features and instance labels, improving voxel representation.
• Enhanced clustering performance without predicting point cloud offsets.

Main Results:

• HyperG-PS demonstrated superior performance on the SemanticKITTI and nuScenes datasets.
• The hypergraph learning module effectively improved instance segmentation clustering accuracy.
• The method achieved state-of-the-art results compared to existing approaches.

Conclusions:

• HyperG-PS offers an effective solution for LiDAR panoptic segmentation by leveraging hypergraph learning.
• The proposed framework significantly improves the understanding of the surrounding environment for autonomous systems.
• This work advances the field of autonomous driving perception.