Multi-Calib: A Scalable LiDAR-Camera Calibration Network for Variable Sensor Configurations.

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Traditional sensor calibration is manual and time-consuming.
• Existing automated methods struggle with scalability and computational demands for multiple sensors.

Purpose of the Study:

• To develop a lightweight, scalable network for joint calibration of multiple LiDAR-camera pairs.
• To reduce computational overhead while maintaining high calibration accuracy.

Main Methods:

• Utilized a frozen pre-trained Swin Transformer for unified feature extraction from RGB images and depth maps.
• Introduced a cross-modal channel-wise attention module for feature alignment and noise reduction.
• Designed a modular calibration head for independent extrinsic estimation per sensor pair.

Main Results:

• Achieved comparable performance to existing methods on the nuScenes dataset.
• Demonstrated a mean translation error of 2.651 cm and rotation error of 0.246 degrees.
• Model requires only 78.79 M parameters, significantly reducing computational cost.

Conclusions:

• The proposed lightweight network offers an efficient and scalable solution for multi-sensor calibration.
• The method effectively handles viewpoint variations and achieves high accuracy.
• This approach is suitable for large-scale deployment in autonomous systems.