UnVELO: Unsupervised Vision-Enhanced LiDAR Odometry with Online Correction

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Visual and LiDAR data offer complementary information for vision tasks.
• Existing visual-LiDAR odometries (VLOs) are under-explored, with most focusing on single modalities.

Purpose of the Study:

• To propose a novel unsupervised visual-LiDAR odometry (VLO) method.
• To develop a LiDAR-dominant fusion scheme for enhanced VLO performance.

Main Methods:

• Implemented an unsupervised vision-enhanced LiDAR odometry (UnVELO) using a LiDAR-dominant approach.
• Fused 3D LiDAR points (spherical projection) with visual information to create dense vertex and color maps.
• Utilized point-to-plane geometric loss and photometric visual loss, incorporating an online pose-correction module.

Main Results:

• UnVELO demonstrated superior performance compared to previous two-frame learning methods on KITTI and DSEC datasets.
• The LiDAR-dominant fusion approach improved robustness to illumination variations and online pose correction efficiency.
• Achieved competitive results against hybrid methods integrating global optimization.

Conclusions:

• The proposed UnVELO method offers an effective approach to unsupervised visual-LiDAR odometry.
• LiDAR-dominant fusion with dense representations enhances visual-LiDAR integration and robustness.
• UnVELO presents a promising direction for advancing autonomous navigation systems.