Roadside Camera-LiDAR Calibration Without Annotation

Applying laser detection and ranging (LiDAR) and camera fusion sensing in roadside environments enhances traffic information accuracy compared to single-sensor systems. Achieving precise spatial calibration is crucial for high-precision fused sensing. Traditional spatial calibration methods are often time-consuming and labor-intensive or necessitate the use of ground truth data for training which is difficult to obtain. Therefore, we introduce an online LiDAR and camera calibration method named roadside online and initialization calibration (ROIC). Unlike prior research, ROIC eliminates the need for ground truth labeling and solely relies on input from point cloud sequences and camera data to complete the spatial calibration of LiDAR and camera. Meanwhile, ROIC reduces the investment of manpower and time and improves the calibration efficiency while ensuring the calibration accuracy. ROIC follows a three-step process. First, preprocessing and semantic segmentation of point clouds and images are performed to get the required class information. Next, the initial extrinsic parameters are estimated using the object centroids. Finally, the extrinsic parameter is optimized using the defined loss function. To evaluate the effectiveness of our approach, we conducted tests using the DAIR dataset. The results demonstrate that the ROIC method achieves a reprojection error of 11.1715, which is superior to the official calibration parameters' reprojection error of 17.746.