TL;DR
This paper introduces a novel LiDAR point cloud map coding scheme that treats maps as sequences, leveraging segmentation-based intra-prediction and an interpolation network for inter-frame coding. This approach achieves significant compression ratios (up to 1/24th of original size) while maintaining high precision (2mm-level), outperforming existing methods like Octree and Google Draco.
What Problem Does This Paper Address?
LiDAR point clouds are crucial for applications like autonomous driving, enabling tasks such as obstacle detection and navigation. However, these dense 3D maps are massive, requiring substantial storage and bandwidth. Existing compression methods, like Octree-based approaches, often struggle to balance compression efficiency with the high precision demands of critical applications. Furthermore, they don't fully exploit the inherent redundancies within point cloud maps, which are essentially sequences of scans. This paper aims to bridge this gap by developing an efficient coding scheme that effectively removes both spatial and temporal redundancies in LiDAR point cloud maps.
Methods
The proposed scheme frames LiDAR point cloud map compression as a point cloud sequence compression problem. This is achieved by recording the time stamp and quaternion matrix of each scan during map building. The coding architecture comprises two main techniques:
- Intra-coding: For individual frames (I-frames), a segmentation-based intra-prediction technique is employed. This involves:
- Converting the 3D point cloud to a 2D representation.
- Using the RangeNet++ network for semantic segmentation to identify different objects and ground.
- Fitting quadric surfaces (planes for ground, spheres for objects) to the segmented regions to predict point cloud data.
- Encoding the residual data (difference between original and predicted data) using lossless coding.
- Inter-coding: For frames that depend on others (B-frames), an interpolation-based inter-frame coding network is used to exploit temporal redundancy. This network:
- Takes two previously decoded frames as input.
- Generates a virtual point cloud for the intermediate frame through interpolation.
- Encodes the residual data between the actual and predicted intermediate frame.
The final point cloud map is reconstructed using the decoded point cloud sequences and their associated time stamps and quaternion matrices.
Datasets
The paper utilizes the KITTI dataset for evaluation. This dataset includes various scenes such as campus, city, residential, and road, providing a realistic benchmark for autonomous driving scenarios. No new benchmarks were introduced.
Key Insights and Ideas
- Treating maps as sequences is key: By leveraging the temporal nature of LiDAR scans and their associated metadata (time stamps, quaternions), the problem can be effectively reframed as point cloud sequence compression, unlocking new avenues for redundancy reduction.
- Segmentation unlocks efficient prediction: Semantic segmentation of point clouds allows for more accurate and context-aware prediction. Fitting geometric primitives (planes, spheres) to these segmented regions is more effective than generic prediction methods.
- Interpolation network bridges temporal gaps: The proposed interpolation network effectively synthesizes intermediate frames, significantly reducing temporal redundancy and improving compression efficiency.
- Residual coding is crucial: Encoding only the difference between the original and predicted data (residual) is a standard but highly effective technique for compression, especially when the prediction is accurate.
Key Results
- High Compression Ratio: The proposed scheme achieves a compression ratio of up to 1/24th of the original size.
- High Precision: The coding maintains a 2 mm-level precision.
- Performance Comparison: The algorithm demonstrates better coding performance compared to established methods like Octree and Google Draco across various scenes. For instance, the average compression ratio achieved by the proposed method was 3.61% (2mm QA), significantly outperforming Octree (average 20.23%) and Draco (average 6.22% for 14-bit QB).
- Qualitative Results: Visualizations show that the generated point clouds closely resemble the ground truth, indicating effective reconstruction.
Limitations and Future Work
- Real-time Performance: The paper acknowledges that the current intra-coding and inter-coding methods do not meet real-time requirements. The average intra-coding time was 0.21s and inter-coding time was 0.15s per frame.
- Future Work: The authors plan to focus on implementing the coding scheme on an FPGA platform to accelerate the algorithm and achieve real-time performance. This would make the scheme more suitable for live autonomous driving applications.
Summary updated