Simple python script for projecting a camera pixel into the ground plane. Require full camera calibration (intrinsic/extrinsic parameters)

project_pixel_to_floor.py

#!/usr/bin/env python 

"""project_pixel_to_floor.py project an image pixel into the ground plane"""

__author__ = "Eric Cox"
__maintainer__ = "Eric Cox"
__email__ = "[email protected]"

import numpy as np
from numpy.linalg import inv
from tf import transformations as tfx


def project_camera_point_to_ground_plane(pixel_coordinate,
                                         camera_position,
                                         camera_orientation,
                                         camera_intrinsics):
    """ Inputs:
    pixel_coordinate: list [u, v] containing pixel coordinate
    camera_position: list [x, y, z] describing camera origin
    camera_orientaion: list [qx, qy, qz, qw] describing camera orientation
    camera_intrinsics: list of 9 camera parameters

    Outputs:
    floor_point: list [x, y, z] containing pixel projection in robot frame
    """
    # Find ground plane normal in camera frame
    rotation = np.matrix(tfx.quaternion_matrix(camera_orientation))[0:3, 0:3]
    plane_normal_in_robot_frame = np.matrix([[0.], [0.], [camera_position[2]]])
    plane_normal = -rotation.T * plane_normal_in_robot_frame

    # Solve for unit vector describing pixel location in camera frame
    # TODO: If distortion is known, we must also take it into account
    camera_vector = inv(camera_intrinsics) * np.matrix(
        pixel_coordinate + [1.]).T

    # find scaling factor that projects normalized camera ray to floor plane
    # TODO: avoid div0. Note that div0 only occurs when unit vector is
    # parallel to ground plane. In this case, it is silly to try to project it
    # into the ground plane to begin with... ¯\_(ツ)_/¯
    sf = camera_position[2] ** 2 / np.dot(camera_vector.T.tolist()[0],
                                          plane_normal.T.tolist()[0])
    camera_vector = sf * camera_vector

    # Convert camera_vector to robot_frame
    floor_point = np.matrix(camera_position).T + rotation * camera_vector

    if np.abs(floor_point[2]) > 1e-6:
        raise Exception("Point is not in the floor plane! Eric is dumb!")
    else:
        # Clear out any numeric error
        floor_point[2] = 0.

    return floor_point.T.tolist()[0]

if __name__ == "__main__":

    # pose of camera with respect to base_footprint (could also be odom)
    # `rosrun tf tf_echo /base_footprint /depthcam_proximity_front_depth_optical_frame`
    quaternion = [0.703, 0.699, 0.095, 0.092]
    translation = [0.109, -0.010, 1.813]

    # Camera intrinsics are determined experimentally for each camera
    # (This is given in the camera info topic)
    K = np.matrix([[142.5851287841797, 0.0, 79.25],
                   [0.0, 142.5851287841797, 56.0],
                   [0.0, 0.0, 1.0]])

    # Pixel coordinates of image boundaries
    pts = [[0, 0], [0, 120], [160, 0], [160, 120]]
    floor_pts = []
    for pt in pts:
        floor_pts.append(project_camera_point_to_ground_plane(pt,
                                                              translation,
                                                              quaternion,
                                                              K))
    print(floor_pts)