ROD2021 Challenge @ ICMR 2021 Forum

Go back to competition Back to thread list Post in this thread

> Is the height of the camera and definition of cooridinate system available?

Dear Mr. YiZhou:
Hi, I want to know if the height of the camera above the ground is available. And could you elaborate on "camera to radar translation:[0.11, -0.05, 0.06]", namely how the coordinate system is defined?
Thanks,
Chuan

Posted by: mrhuangchuan @ March 10, 2021, 1:49 a.m.

The height of the camera may vary in different scenarios. That is the reason why we did not release this parameter.

The coordinates for both camera and radar are defined as follows:
x: right
y: down
z: forward

Posted by: ywang26 @ March 12, 2021, 3:11 a.m.

Okay, but still got a question. I looked at the sensor configure file, it has both projection matrix(P) and camera matrix(K). I tried to recover camera pose by inv(K) * P, it should give me camera pose [R|t], but R matrix(rotation matrix) determinant is not one.
Below are what I did.
-----------------------------------
print('camera intrinsics:')
print(dataset.sensor_cfg.calib_cfg['cam_calib']['2019_04_09']['cam_0']['camera_matrix'])
print('camera to radar translation:')
print(dataset.sensor_cfg.calib_cfg['t_cl2rh'])
import numpy as np
K = np.asarray(dataset.sensor_cfg.calib_cfg['cam_calib']['2019_04_09']['cam_0']['camera_matrix'])
P = np.asarray(dataset.sensor_cfg.calib_cfg['cam_calib']['2019_04_09']['cam_0']['projection_matrix'])
T = np.linalg.inv(K) @ P
print(T)
--------------------output------------------------------------
camera intrinsics:
[[849.177455 0. 712.166787]
[ 0. 854.207389 543.445028]
[ 0. 0. 1. ]]
camera to radar translation:
[0.11, -0.05, 0.06]
[[0.74903762 0. 0.0083702 0. ]
[0. 0.8322183 0.00211336 0. ]
[0. 0. 1. 0. ]]

Posted by: mrhuangchuan @ March 12, 2021, 3:20 a.m.

The camera poses cannot be recovered using the provided camera calibration results. The provided calibration parameters are based on a checkerboard using OpenCV (https://docs.opencv.org/2.4/modules/calib3d/doc/camera_calibration_and_3d_reconstruction.html). It is not clear what you are trying to obtain. Since camera poses are always different for our moving camera scenario, you can try to obtain the camera pose for each frame using visual odometry.

Posted by: ywang26 @ March 12, 2021, 3:39 a.m.
Post in this thread