CAE-LO: LiDAR Odometry Leveraging Fully Unsupervised Convolutional Auto-Encoder for Interest Point Detection and Feature Description

Deyu Yin, Qian Zhang, Jingbin Liu, Xinlian Liang, Yunsheng Wang, Jyri Maanpää, Hao Ma, Juha Hyyppä, Ruizhi ChenarXiv 2020 · arXiv

Introduction

LiDAR Odometry (LO) is a fundamental component of autonomous driving systems, enabling vehicles to estimate their motion by analyzing laser scans of their surroundings. Despite its importance, current LO methods face two significant challenges. First, LiDAR point clouds are inherently massive and unordered, making direct processing computationally expensive. Second, many state-of-the-art deep learning approaches rely on supervised training with "ground truth" poses, which are difficult and costly to obtain across diverse real-world environments.

Furthermore, existing methods often struggle in "unstructured" scenarios, such as open highways or rural roads, where large geometric landmarks like building walls are absent. These systems frequently overlook fine-grained local details that are essential for stable tracking when prominent landmarks are missing. CAE-LO addresses these issues by introducing a fully unsupervised framework that utilizes Convolutional Auto-Encoders (CAEs) for both interest point detection and feature description.

Methodology

System Overview

The CAE-LO pipeline is designed to be a hybrid system that leverages the efficiency of 2D representations for detection and the geometric richness of 3D representations for description. The process begins with raw point cloud sequences and concludes with a refined final odometry estimate.

The overall CAE-LO pipeline, illustrating the flow from raw data to final odometry through interest point detection, feature extraction, and pose refinement.

2D-CAE Interest Point Detection

To handle the computational complexity of 3D data, the system first projects the raw 3D point cloud into a compact 2D spherical ring image. A 2D Convolutional Auto-Encoder is trained in an unsupervised manner to compress and reconstruct these images. Interest points are identified by analyzing the learned feature vectors; pixels that differ significantly from their neighbors are flagged as points of interest. This approach is computationally efficient and excels at capturing fine local details that traditional geometric methods might miss.

3D-CAE Multi-Scale Feature Description

Once interest points are identified, the system shifts to a 3D representation to ensure robust description. Voxel patches are extracted around each interest point at three distinct resolutions: small, mid, and big scales.

The multi-scale voxel model extracts features at three different scales to capture both high-precision local geometry and broader spatial context.

These patches are processed by a 3D-CAE that compresses them into a 60-dimensional descriptor. This "pyramid" approach allows the system to maintain high-precision local geometry while simultaneously considering the broader spatial context, leading to more stable matching across frames.

Odometry Refinement and Pose Update

The initial motion estimate is generated using a RANSAC-based approach, followed by an Iterative Closest Point (ICP) refinement on keyframes. To ensure a smooth trajectory, the authors implement a "Backward Pose Update." This method prevents sudden jumps in the estimated path when a keyframe is refined by distributing the pose correction back to the ordinary frames located between keyframes.

Key Insights

  • Hybrid Representation: The core innovation lies in using 2D grids for rapid point detection and 3D voxels for accurate shape description. This "best of both worlds" approach optimizes both speed and precision.
  • Unsupervised Universality: Because the CAEs are trained solely to reconstruct their own input, the model requires no human-labeled data. This makes CAE-LO highly adaptable to different LiDAR sensors and environments without the need for expensive retraining.
  • Robustness in Sparse Environments: Unlike methods that rely on large structures, CAE-LO’s sensitivity to local feature differences allows it to find stable tracking points (like small corners or road edges) even on open highways where other algorithms often fail.

Results and Performance

The framework was evaluated using the KITTI Odometry Benchmark, comparing it against leading unsupervised and weakly-supervised methods such as USIP and 3DFeat-Net.

Key performance metrics include:

  • Matching Accuracy: CAE-LO achieved a matching inlier ratio more than 50% higher than previous state-of-the-art methods.
  • Success Rate: The system demonstrated a 99.8% success rate on the KITTI dataset, outperforming 3DFeat-Net (99.5%) and USIP (99.5%).
  • Error Reduction: The refinement process and backward pose update significantly improved accuracy, reducing relative translation error from 0.354m to 0.170m.

Conclusion

CAE-LO represents a significant step forward in unsupervised LiDAR odometry. By combining 2D and 3D deep learning techniques, it provides a robust, label-free solution that excels in challenging, unstructured environments. Future work will focus on improving rotation invariance in descriptors and exploring the use of spherical ring representations for additional tasks like semantic segmentation.

Summary updated

BibTeX
@misc{yin2020cae,
  author = {Deyu Yin and Qian Zhang and Jingbin Liu and Xinlian Liang and Yunsheng Wang and Jyri Maanpää and Hao Ma and Juha Hyyppä and Ruizhi Chen},
  title = {{CAE-LO: LiDAR Odometry Leveraging Fully Unsupervised Convolutional Auto-Encoder for Interest Point Detection and Feature Description}},
  year = {2020},
  eprint = {2001.01354},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2001.01354},
}
Figure