Projecting Points with ROS Pinhole Camera Model

A common task in robotics-related Computer Vision projects is projecting a 3D point into the 2D image plane of a camera, or vice-versa. In ROS this can be achieved very easily with the help of classes tf.TransformListener and image_geometry.PinholeCameraModel.

Given an arbitrary 3D point in world coordinates, the first step is to convert its coordinates from the global reference frame to a camera-centric frame. Given a tuple p_3d containing global coordinates of a point, and assuming there is a tf broadcaster publishing transforms between the /world (global) and /base_link (camera-centric) frames, the code below transforms p_3d to camera-centric coordinates:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
p_world = PointStamped()
p_world.header.seq = self.camera_image.header.seq
p_world.header.stamp = stamp
p_world.header.frame_id = '/world'
p_world.point.x = p_3d[0]
p_world.point.y = p_3d[1]
p_world.point.z = p_3d[2]

listener = tf.TransformListener()
listener.waitForTransform('/base_link', '/world', stamp, rospy.Duration(1.0))
p_camera = listener.transformPoint('/base_link', p_world)

The camera-centric point p_camera can then be projected into the image as in the code below:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
# The navigation frame has X pointing forward, Y left and Z up, whereas the
# vision frame has X pointing right, Y down and Z forward; hence the need to
# reassign axes here.
x = -p_camera.point.y
y = -p_camera.point.z
z = p_camera.point.x

camera = PinholeCameraModel()
camera.fromCameraInfo(camera_info)
p_image = camera.project3dToPixel((x, y, z))

The camera_info argument is an instance of the sensor_msgs.msg.CameraInfo message class, it can be instantiated locally or received through a topic.

Notice that in real-world applications, rather than locally instantiating tf.TransformListener and image_geometry.PinholeCameraModel objects as in the code above, they'd more likely be instantiated once and kept as attributes of a node object.

Pages