Aerospace and Electronic Systems - November 2018 - 24

Optical Flow Navigation System for a Flying Robot
min bias drift in x- and ydirections and 0.4 deg/min
bias drift in the z-direction.
Moreover, the update rate
is 780 Hz and the maximum measurable angular
rate is 500 °/s.
In order to implement a
real-time BM, as explained
previously, we first divided
each frame into blocks
with dimensions of 8 x 8
pixels and also adopted
the related search windows
with dimensions of ±4 pixels. Subsequently, we used
the Harris Corner Detector
(HCD) [29] to detect salient points and examined Figure 3.
the suitability of pixel lo- Px4flow sensor, Pixhawk autopilot, and the developed vehicle.
cations for OF reckonings.
To indicate the best match for the detected salient points in the
previous step, the SAD for the adopted block in the current frame
and every block in the previous frame is computed. Then the motion vector (u, v), which stands for the displacement of a block
in two consecutive frames, in x- and y- directions (Figure 1) is
determined. It should be noted that the precision in vision-based
navigation systems depends heavily on the quality of images captured by the vision sensor [30]. At this phase, to show the effects of
the illumination condition and image quality on the performance of
the optical navigation system, we defined a quality parameter/factor. This factor is calculated based on the number of pixel locations
which are suitable for flow estimation as noted below:
Q=

255 × j
k×k

(20)

in which Q is the image quality and j is the number of tiles to
check. If there is any suitable pixel location, based on HCD, the
corner detector uses minimum eigenvalues [29] that act as threshold values for the quality level of a corner. If a corner is below
the quality threshold value, that corner is not tracked. In (20), k
indicates the number of corners which are tracked. In fact, this parameter shows the effect of illumination on the precision of optical
flow calculations and is quantified from 0 to 255. Zero value is
acquired when there is no suitable pixel location under the worst illumination condition. On the other hand, the quality factor reaches
255 when all pixel locations in the search window are suitable for
OF calculations under a perfect illumination condition.

TEST RESULTS
As stated before, in order to evaluate the performance of the mentioned three optical navigation methods (LK, HS, and BM), an
off-line implementation was conducted on captured images of the
vision sensor mounted on an in-house-built quadrotor. Since the
24

Figure 4.

Components of pixel-based velocity estimated by LK, HS, and BM
methods.

quality of the images and the degree of the precision for navigation
purposes depend on the vehicle's orientation during the flight [5],
in this research we tested the proposed optical navigation system at
almost constant altitudes.
Based on the test scenarios, the vehicle was flown over the Kooye
Danesgah area at the University of Tehran Campus. Both x and y
components of the pixel velocity are calculated through LK, HS, and
BM as explained previously. In each frame, every pixel's velocity is
estimated and then the average of the values is assumed as the image
velocity (Figure 4). As far as this research is concerned, we compared
the precision of the three methods in terms of standard deviation
of pixel velocity, estimated at each pixel location of the considered
frame (Figure 5). Since low dispersion is associated with high precision, the lower the data dispersion, the more precise the estimation.
Subsequently, as illustrated in Figure 5, since the dispersion of pixel

IEEE A&E SYSTEMS MAGAZINE

NOVEMBER 2018

Aerospace and Electronic Systems - November 2018

Table of Contents for the Digital Edition of Aerospace and Electronic Systems - November 2018