LiDAR-camera system extrinsic calibration by establishing virtual point correspondences from pseudo calibration...

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Extrinsic calibration of LiDAR-camera systems is crucial for sensor fusion but challenging without specific calibration objects.
• Finding accurate point correspondences between sparse LiDAR data and dense RGB images is a significant hurdle.

Purpose of the Study:

• To propose a novel LiDAR-camera extrinsic calibration method that does not require specific calibration objects.
• To introduce the concept of virtual point correspondences for robust calibration.

Main Methods:

• Utilizing pseudo calibration objects in natural scenes that meet specific criteria.
• Establishing virtual point correspondences from 2D image bounding boxes and 3D LiDAR frustum boxes.
• Employing a normal-guided foreground detection for automatic 3D frustum box extraction.
• Applying a geometrical optimization scheme to estimate extrinsic parameters.

Main Results:

• The proposed method achieves accurate and robust LiDAR-camera extrinsic calibration.
• Performance is validated through simulations and real-world experiments.
• Outperforms existing calibration object-based methods.

Conclusions:

• The novel virtual point correspondence approach enables effective LiDAR-camera extrinsic calibration without specialized objects.
• The method is practical due to easily satisfied calibration conditions.
• Demonstrates significant advancements in autonomous system calibration.