SDV-LOAM: Semi-Direct Visual-LiDAR Odometry and Mapping.

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Sensor Fusion

Background:

• Visual-LiDAR odometry and mapping (V-LOAM) systems fuse camera and LiDAR data for accurate pose estimation.
• Existing V-LOAM systems face challenges with 3D-2D depth association and vertical drift in 6-DOF optimization.
• These limitations lead to nontrivial tracking errors in pose estimation and mapping.

Purpose of the Study:

• To present SDV-LOAM, a novel system designed to overcome the tracking errors in V-LOAM systems.
• To achieve high tracking accuracy and robustness in pose estimation and mapping.
• To improve the performance of both visual and LiDAR odometry components.

Main Methods:

• Incorporates a semi-direct visual odometry (VO) module that tracks high-gradient pixels where 3D LiDAR points project.
• Employs a novel point matching with propagation method to mitigate scale differences in VO.
• Utilizes an adaptive sweep-to-map LiDAR odometry that dynamically selects between 3-DOF and 6-DOF optimization based on vertical geometric constraints.
• Introduces a sweep reconstruction method to synchronize LiDAR point cloud input frequency with camera images.

Main Results:

• SDV-LOAM achieved 8th place on the KITTI odometry benchmark, outperforming most LiDAR/visual-LiDAR systems.
• The visual module surpassed state-of-the-art VO systems in performance.
• The adaptive sweep-to-map optimization enhanced existing open-source LiDAR odometry systems.
• Demonstrated high accuracy and output frequency on a custom hardware platform in large-scale environments.

Conclusions:

• SDV-LOAM effectively addresses critical error sources in V-LOAM, significantly improving tracking accuracy.
• The system demonstrates superior performance compared to existing methods on benchmark datasets and real-world applications.
• The source code is released to foster community development in visual-LiDAR odometry and mapping.