Efficient 3D Deep LiDAR Odometry (2111.02135v2)

Abstract: An efficient 3D point cloud learning architecture, named EfficientLO-Net, for LiDAR odometry is first proposed in this paper. In this architecture, the projection-aware representation of the 3D point cloud is proposed to organize the raw 3D point cloud into an ordered data form to achieve efficiency. The Pyramid, Warping, and Cost volume (PWC) structure for the LiDAR odometry task is built to estimate and refine the pose in a coarse-to-fine approach. A projection-aware attentive cost volume is built to directly associate two discrete point clouds and obtain embedding motion patterns. Then, a trainable embedding mask is proposed to weigh the local motion patterns to regress the overall pose and filter outlier points. The trainable pose warp-refinement module is iteratively used with embedding mask optimized hierarchically to make the pose estimation more robust for outliers. The entire architecture is holistically optimized end-to-end to achieve adaptive learning of cost volume and mask, and all operations involving point cloud sampling and grouping are accelerated by projection-aware 3D feature learning methods. The superior performance and effectiveness of our LiDAR odometry architecture are demonstrated on KITTI, M2DGR, and Argoverse datasets. Our method outperforms all recent learning-based methods and even the geometry-based approach, LOAM with mapping optimization, on most sequences of KITTI odometry dataset. We open sourced our codes at: https://github.com/IRMVLab/EfficientLO-Net.

• The paper introduces EfficientLO-Net, a projection-aware deep learning framework that enhances LiDAR odometry accuracy via an end-to-end Siamese architecture.
• It employs a Siamese Point Feature Pyramid, a Projection-Aware Attentive Cost Volume, and iterative pose warp-refinement to optimize pose estimates.
• The approach achieves 20 Hz real-time processing on standard datasets, outperforming both traditional and learning-based methods.

Efficient 3D Deep LiDAR Odometry

This paper introduces a novel approach for performing LiDAR odometry using 3D point clouds, named EfficientLO-Net. The authors propose a new architecture designed to improve the computational efficiency and accuracy of LiDAR odometry by leveraging end-to-end deep learning techniques.

Key Contributions

1. Projection-Aware Representation: The paper utilizes a projection-aware approach to handle 3D point cloud data, organizing it efficiently in a cylindrical projection space. This representation retains the raw 3D point clouds while making use of their line scan characteristics to improve computational efficiency in subsequent processing.
2. Network Architecture: The EfficientLO-Net architecture consists of a Siamese Point Feature Pyramid, a novel Projection-Aware Attentive Cost Volume, and an iterative Pose Warp-Refinement module. This architecture enables the network to refine pose estimates through a coarse-to-fine methodology.
3. Hierarchical Embedding Mask Optimization: A trainable embedding mask is incorporated to weigh the local motion patterns within point clouds, enhancing robustness to outliers and dynamically moving objects. The mask is refined in a hierarchical fashion to improve accuracy in the final pose estimation.
4. Comprehensive Evaluation: The authors provide extensive evaluations on well-known datasets, including KITTI, M2DGR, and Argoverse, demonstrating superior performance compared to both learning-based and traditional geometry-based methods. Notably, EfficientLO-Net outperformed existing state-of-the-art methods in several sequences in the KITTI odometry dataset.

Results and Implications

EfficientLO-Net achieves remarkable performance, particularly in terms of real-time processing capability, yielding results at 20 Hz while maintaining high accuracy. Its end-to-end optimization of odometry and mask filtering contributes to the resilience against typical challenges in LiDAR data, such as occlusion and dynamic environments. Furthermore, EfficientLO-Net offers computational advantages, significantly reducing inference time compared to previous methods.

Discussion and Future Work

The approach not only achieves superior results compared to conventional methods like LOAM but also demonstrates the potential of deep learning in handling large-scale point cloud data more effectively. The integration of 3D learning techniques with efficient data structuring heralds a significant step forward in LiDAR odometry applications.

Looking to the future, this framework could be extended to enhance other real-time LiDAR-based tasks, such as large-scale semantic segmentation and 3D object detection. Moreover, by integrating with mapping systems, it may also serve as a robust front-end odometry solution. Researchers will likely build upon this work to explore further applications and optimizations within the scope of autonomous navigation and robotic perception.