Inter-LPCM: Learning-Based Inter-Frame Predictive Coding for LiDAR Point Cloud Compression

Area of Science:

• Computer Vision
• Signal Processing
• Data Compression

Background:

• LiDAR point clouds are efficiently compressed using spherical coordinates.
• Existing methods like PredGeom have limitations in capturing complex motion and inter-frame correlations.
• Learning-based methods in the spherical domain often neglect inter-frame redundancy.

Purpose of the Study:

• To develop a learning-based inter-frame predictive coding method (Inter-LPCM) for LiDAR point clouds.
• To enhance compression efficiency by exploiting inter-frame correlations in the spherical domain.
• To improve upon the rate-distortion performance of existing point cloud compression techniques.

Main Methods:

• Proposed Inter-LPCM utilizes delta coding for azimuth prediction.
• Introduced an inter-frame radius predictive (Inter-RP) model for radius compression.
• Designed a lightweight attention-based prediction (LAEP) model for elevation angle prediction.
• Implemented RD-optimized quantization and distinct entropy coding models for spherical components.

Main Results:

• Inter-LPCM achieved a 26.1% BD-rate reduction compared to G-PCC lossless octree coding on SemanticKITTI.
• Demonstrated an 8.3% BD-rate reduction compared to PredGeom's inter-frame mode on Ford dataset.
• The method effectively captures long-range geometric correlations and exploits inter-frame redundancies.

Conclusions:

• Inter-LPCM offers superior rate-distortion performance for LiDAR point cloud compression.
• The proposed method advances the state-of-the-art in spherical domain point cloud compression.
• Publicly available source code facilitates further research and development.