Deep Learning-Based Lightweight LiDAR and Fisheye Camera Online Extrinsic Calibration

Autonomous driving has been extensively researched in recent years. To improve the mobility and decision-making capabilities of autonomous vehicles, multiple sensors have been integrated to complement the limitations of individual sensors. For example, Light Detection and Ranging (LiDAR) is frequently combined with camera data to overcome the narrow field-of-view (FoV) of traditional pinhole cameras. The fisheye cameras of LiDAR expand the FoV from up to 80° in traditional cameras to 180°, which is advantageous for autonomous driving applications. This study introduces a lightweight, deep learning-based LiDAR-fisheye camera fusion model for real-world environments. The mean translation errors are 1.375, 0.753, and 1.208 cm along the X, Y, and Z axes, respectively, and the mean rotation errors are 0.171°, 0.150°, and 0.089° in the roll, pitch, and yaw directions, respectively. These results demonstrate the efficiency and proficiency of our sensor-fusion approach for autonomous driving.