A novel 3D LiDAR deep learning approach for uncrewed vehicle odometry

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Precise self-localization and pose registration are critical for autonomous vehicle operation in unpredictable environments.
• Existing methods like Convolutional Neural Network (CNN)-based approaches may lose information due to cylindrical projection preprocessing.
• Accurate localization and mapping are fundamental for odometry, path planning, and other essential autonomous driving functions.

Purpose of the Study:

• To develop an advanced LiDAR-based localization and mapping (LOAM) technique for autonomous vehicles.
• To mitigate information loss inherent in preprocessing steps of traditional methods.
• To improve the accuracy and robustness of pose estimation for autonomous navigation.

Main Methods:

• Leveraging point cloud-based deep learning for LiDAR-based localization and mapping (LOAM), avoiding cylindrical projection required by CNNs.
• Employing the Normal Distribution Transform (NDT) algorithm to refine coarse pose estimations obtained from the deep learning model.
• Investigating the use of Product Quantization with high-level features as fingerprints to optimize NDT's internal neighborhood searching.

Main Results:

• The proposed deep learning and NDT refinement method achieves performance comparable to state-of-the-art benchmark studies.
• The point cloud-based deep learning approach effectively reduces information loss compared to projection-based methods.
• Product Quantization shows potential for enhancing the efficiency of NDT neighborhood searching.

Conclusions:

• The integrated deep learning and NDT approach offers a robust and accurate solution for autonomous vehicle localization and mapping.
• This method provides a viable alternative to traditional techniques, particularly in challenging, uncertain environments.
• Further research into feature-based optimization of NDT could lead to significant improvements in real-time performance.