Navigation Functionalities for an Autonomous UAV Helicopter
Navigation Functionalities for an Autonomous UAV Helicopter
Navigation Functionalities for an Autonomous UAV Helicopter
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
A.3. PAPER III 99<br />
the projection of three circles lying on the corner points of <strong>an</strong> equilateral<br />
tri<strong>an</strong>gle the pose of <strong>an</strong> object is uniquely determined, assuming all intrinsic<br />
camera parameters are known. Circles are projected as ellipses, described<br />
by the center point ue, the semi-major axis la, the semi-minor axis lb, <strong>an</strong>d<br />
the semi-major axis <strong>an</strong>gle θe The pose of the l<strong>an</strong>ding pad with respect to<br />
the camera coordinate system is estimated by minimizing the reprojection<br />
error of the extracted center points <strong>an</strong>d semi-axes of the three ellipses. We<br />
use five circle triplets of different size (radius 2 to 32 cm, dist<strong>an</strong>ce 8 to 128<br />
cm) with common center point to achieve a wide r<strong>an</strong>ge of possible camera<br />
positions. Each triplet is uniquely determined by a combination of differently<br />
sized inner circles.<br />
A point p ˜x in the l<strong>an</strong>ding pad frame is projected on the image pl<strong>an</strong>e as<br />
follows:<br />
ũ = P p ⎛ ⎞<br />
αu 0 u0 0 �<br />
cpR<br />
˜x = ⎝ 0 αv v0 0 ⎠<br />
0 0 1 0<br />
ctp 0T �<br />
p˜x 2 p 3<br />
ũ ∈ P ˜x ∈ P<br />
3 1<br />
The extrinsic camera parameters are given by the three Euler <strong>an</strong>gles of the<br />
rotation matrix c pR <strong>an</strong>d the three components of the tr<strong>an</strong>slation vector c tp. We<br />
use a camera model with the following intrinsic parameters: ”focal lengths”<br />
αu <strong>an</strong>d αv in pixels, principal point (u0, v0), <strong>an</strong>d four lens distortion coefficients<br />
. All intrinsic parameters are calibrated using Bouguet’s calibration<br />
toolbox [3]. A conic in P 2 is the locus of all points ũ satisfying the homogeneous<br />
quadratic equation ũ T C ũ = 0. The tr<strong>an</strong>s<strong>for</strong>mation of a circle Cp on<br />
the l<strong>an</strong>ding pad into <strong>an</strong> ellipse Ci in the image pl<strong>an</strong>e is given by[4]:<br />
Ci = (H −1 ) T Cp H −1<br />
The homography matrix H is the projection matrix P without third column<br />
(z = 0). We calculate the ellipse center <strong>an</strong>d axes from Ci <strong>an</strong>d represent the<br />
parameters in a common feature vector c.<br />
Fig. 3 shows a data flow diagram of the vision system. Round-edged boxes<br />
represent image processing functions, sharp-edged boxes indicate independent<br />
processes, <strong>an</strong>d dashed lines show trigger connections. Closed contours are<br />
extracted from gray-level images using a fast contour following algorithm<br />
with two parameters: edge strength <strong>an</strong>d binarization threshold. The latter<br />
is calculated from the intensity distribution of the reference pattern. In the<br />
contour list we search <strong>for</strong> the three biggest ellipses belonging to a circle<br />
triplet. Ellipse parameters are estimated by minimizing the algebraic dist<strong>an</strong>ce<br />
of undistorted contour points to the conic using SVD [4,6]. After having<br />
found three ellipses, the corresponding contours are resampled with sub-pixel<br />
accuracy. A coarse pose is estimated based on the ratio of semi-major axes<br />
(1)<br />
(2)