Laser-Camera Calibration Questions

[af-doom]

Hello, thank you for your open source dataset. My research direction is multi-sensor calibration. I calibrated the external parameters of the laser and left camera(/stereo/vehicle_frame_left/image_raw/compressed) on the sequence of the vehicle platform. However, in terms of external rotation, my results always differ from the reference value by 1°. In terms of translation, with the laser as the reference, in my obtained results, y parameter is always larger than 34cm, closer to 38. In particular, the Z axis parameter is always larger than -3cm, closer to -8cm. Because the translation obtained by my method on different sequences fluctuates, which is common in all target-free calibration methods. However, the rotation I obtained is extremely stable, but it always differs from the reference value by one degree. I would like to know what the installation parameters of the sensor are on the physical platform, and I would like to compare them. I always believe that the calibration board based approach is very robust, but the targetless approach uses more data. I believe this is more beneficial for external rotation calibration. Finally, I give my calibration results on the two sequences The following are the reference values ​​in the link, converted according to the quaternion [x, y, z, w] [ -0.0175979, -0.0007485, 0.9998448; 0.3502 -0.9998432, 0.0019689, -0.0175964; -0.0332 -0.0019554, -0.9999978, -0.0007830 -0.0332 ]

[gogojjh]

Thanks for your detailed investigation of the extrinsics. To be honest, the installization of the camera on the vehicle is not accurately designed by CAD. The only thing I can provide the rouge baseline (as shown here)

Since extrinsics were calibrated before data collection (before 2 days), the installation may potentially affect the extrinsics. I think that the extrinsics you calibrated may high accuracy. Do you have any methods to evaluate the accuracy of your calibrated extrinsics against ours?

[af-doom]

感谢您对外在因素的详细调查。老实说，摄像头在车辆上的安装并不是由CAD准确设计的。我唯一能提供的胭脂基线（如图所示）

由于外在函数在数据收集之前（2 天之前）进行了校准，因此安装可能会影响外函数。我认为您校准的外在因素可能具有很高的准确性。你有什么方法可以评估你校准的外在因素与我们的精度吗？

Thank you very much for your reply

[af-doom]

感谢您对外在因素的详细调查。老实说，摄像头在车辆上的安装并不是由CAD准确设计的。我唯一能提供的胭脂基线（如图所示） 由于外在函数在数据收集之前（2 天之前）进行了校准，因此安装可能会影响外函数。我认为您校准的外在因素可能具有很高的准确性。你有什么方法可以评估你校准的外在因素与我们的精度吗？

Thank you very much for your reply. To be honest, I think that the targetless calibration is good for the extrinsic rotation calibration, but not accurate enough for the extrinsic translation calibration. I think all the targetless methods are pretty much the same at the moment. Even the almost targetted methods can only achieve accuracy that is consistent with the accuracy of the accurate CAD measurement. As far as the evaluation is concerned, the visual quality evaluation of the projections is almost indistinguishable. I believe that your external translation should be accurate if there is no large mounting deviation, but the external rotation may have errors caused by the mounting deviation. Also, please ask if the sensor is mounted separately on the vehicle, or is the kit rigidly connected during calibration and the kit is directly mounted on the vehicle.