Flow-ICP: semantic segmentation of point clouds based on 4D time-series alignment

Area of Science:

• Computer Vision
• Robotics
• Autonomous Driving

Background:

• Semantic segmentation of LiDAR point clouds is crucial for autonomous systems.
• Temporal information enhances perception in low-visibility or sparse areas.
• Current methods suffer from cumulative errors due to frame-by-frame stacking.

Purpose of the Study:

• To propose a novel method for dynamic alignment of historical LiDAR frames.
• To introduce a new multi-scale feature fusion technique using spatiotemporal (ST) features.
• To improve the accuracy and consistency of semantic segmentation in autonomous driving.

Main Methods:

• Dynamic alignment of historical frame memory to current observations.
• Spatiotemporal (ST) feature extraction for multi-scale feature fusion.
• Optimization and fusion of aligned channel features for enhanced representation.

Main Results:

• The proposed method significantly reduces deviations caused by viewpoint changes and object movements.
• It addresses inconsistencies between 2D range image coordinates and 3D Cartesian outputs.
• Evaluated on SemanticKITTI and SensatUrban datasets, it outperforms state-of-the-art methods.

Conclusions:

• Dynamic alignment and ST feature fusion enhance LiDAR semantic segmentation accuracy.
• The method offers a more robust solution for autonomous driving perception.
• This approach improves feature representation and reduces cumulative errors.