CFNet: LiDAR-Camera Registration Using Calibration Flow Network.

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Sensor Fusion

Background:

• LiDAR-camera calibration is crucial for autonomous vehicle and robot navigation, enabling accurate data fusion.
• Existing calibration methods are often manual, require specific targets, or involve complex, time-consuming targetless optimization.
• Current Convolutional Neural Network (CNN)-based methods for extrinsic parameter regression neglect spatial correspondences, leading to suboptimal performance.

Purpose of the Study:

• To develop a novel, efficient, and accurate CNN-based algorithm for LiDAR-camera extrinsic calibration.
• To overcome the limitations of existing CNN-based methods by explicitly incorporating matching capabilities and geometric information.
• To improve the performance and reduce the complexity of LiDAR-camera calibration for real-world autonomous systems.

Main Methods:

• Proposed CFNet, a CNN-based algorithm utilizing a correlation layer for explicit matching.
• Introduced 'calibration flow' to represent the deviation between projected and ground truth data.
• Employed the efficient Perspective-n-Point (EPnP) algorithm with RANdom SAmple Consensus (RANSAC) to estimate extrinsic parameters using 2D-3D correspondences derived from the calibration flow.

Main Results:

• CFNet demonstrated superior performance compared to state-of-the-art CNN-based methods on the KITTI datasets.
• The method's effectiveness was further validated on the KITTI360 datasets, showcasing its flexibility.
• Explicit consideration of geometric information through calibration flow significantly enhanced accuracy.

Conclusions:

• CFNet offers a more effective and robust approach to LiDAR-camera extrinsic calibration than existing CNN-based techniques.
• The proposed calibration flow mechanism successfully addresses the limitations of previous methods by focusing on spatial correspondences.
• CFNet provides a promising solution for accurate sensor fusion in autonomous driving and robot navigation applications.