![](bg2.png)
optical flow algorithm is employed for local feature
tracking and a local geometric filter (LGF) module
maintaining pairwise geometric consistency is used to
identify outliers.
In all of the above methods (excluding (Viswanath
et al., 2015)), the feature points are identified and
matched between frames, resulting in one extra step
dependent on both the feature matching between
frames algorithm and the feature selection algorithm.
In our method we do not perform any feature match-
ing, using directly the positions estimated by the op-
tical flow as next-frame features and calculate the
homography matrix based on the features found in
the previous frame and the estimated by the optical
flow in the current frame. This allows for a one-
to-one correspondence of features positions between
frames, removing any ambiguity as to which fea-
tures are matched and dramatically enhances perfor-
mance as there is no need for a feature matching step.
In (Viswanath et al., 2015), similar to our method the
feature points in the previous frame are tracked using
an optical flow method and a Homography is calcu-
lated for the background pixels, under the assump-
tion of an initial background state image and subse-
quently updating a distribution of background proba-
bility of each pixels in the new frame. This is used to
model the new background and perform background
subtraction using the entire new background image
as mask, requiring a significant computational bur-
den.This method does not behave well in the absence
of many background points to estimate the Homogra-
phy, and the foreground object estimation needs more
computations, since it operates on the whole image
for both the background transformation/subtraction.
In Section 2 the visual tracking problem of the
UAV from another airborne UAV statement is pro-
vided, followed by the case of identifying a flying
object using the Homography based method in Sec-
tion 2.1. In Section 3 the method of estimating the im-
age based optical flow is discussed using the current
and previous image frames. In Section 4 the method
of estimating the image flow for background image
features based on the Homography transformation de-
rived between image feature points found in previous
and current frames is discussed as well as the method
of comparing the two flows in order to evaluate which
parts of the image are background, that is when the
optical flow velocity vectors match closely those of
the expected by the Homography based flow estima-
tion of the image features corresponding to the back-
ground. The method of calculation of the bounding
box around the tracked target UAV points that have
been identified as foreground is also discussed. In
Section 5 the implementation of the Kalman predic-
tor for the smooth follow of the tracked UAV dur-
ing tracking window estimation noise is discussed. In
Section 6 a correlation based tracker is considered and
described, for further enhancing the ability to track
the object when the flow based method is not pro-
viding a result due to object blending with the back-
ground. In Section 7 the control of the PTZ camera
motion to track the target UAV is discussed along with
the overall algorithm. In Section 8 the results of an
experiment of tracking a UAV from a flying UAV pur-
suer is presented and analyzed.
2 AIRBORNE
VISUAL-TRACKING OF UAV
The algorithms are running locally on a mini i7-based
Intel NUC PC on board of an enhanced Vulkan UAV
shown in Figure 1.
Figure 1: Tracker UAV with mounted PTZ-camera.
Real time LTT of an object using a video-stream is
a complex task due to the limited information we have
about the object. The various correlation methods
give a good tracking performance, given the motion
is not very fast and can keep the correlation strong
between frames, but fail when the motion is fast and
loose the tracking window. These must be initial-
ized with a starting window that engulfs the object
to track. Thus a LTT-scheme based on object motion
is required in order to track a general object across
all motions regions, occlusion and change of appear-
ance. Such a scheme is proposed in this article, en-
hanced with a correlation based tracker for the cases
where the object blends with the background since its
motion which can be slow or not moving relative to it.
2.1 Object Tracking Method using
Homography
The background subtraction method used in order
to identify the moving object motion is relying
only on visual feedback and homography calcula-
Airborne Visual Tracking of UAVs with a Pan-Tilt-Zoom Camera
91