Information Technology Reference
In-Depth Information
6128 and 16GB of RAM was also connected. Given the heterogeneous platform, a
simple load balance method based on the execution time was also implemented.
To denote the number of nodes and number of threads in the graphics, up-
percase letter N and T were used, respectively (e.g. 1N4T for one node and four
threads, 3N24T for three nodes and twenty four threads, etc.).
4.2 Problem Instances
The experimental analysis was performed using the sFly data sets by Lee et
al. [14] ( http://www.sfly.org/mav-datasets ) , which were taken by an UAV
and consist of different images sequences from a front looking and downward
looking cameras, together with measurements from an IMU as well as the ground
truth information given by a precision external reference system called Vicon
system.
The sequence of images used is named hoveringDown ; it consists of 2041
image frames of 752
480 pixels of resolution taken at approximately 20fps
by the downward looking camera. The IMU and Vicon system sample rate are
200Hz, resulting in a total of 21388 samples. They correspond to a flight period
of approximately 106 seconds, where the UAV takes off, performs a hovering
flight and ends landing near to the same place. The total amount of yaw angle
change during the flight is approximately 1 . 4rad.
Furthermore, a data set with own images was also tested. The images were
taken with a 640
×
480 downward looking camera attached to a quadrotor over a
carpet at approximately 15fps. In Fig. 5 two consecutive images of the data set
are shown where it is possible to see the total absence of intensity features. In this
conditions traditional feature trackers are hard to use. The flight period of the
data set is around 40 seconds resulting in a total of 600 images without features.
The quadrotor is named QA3 and it is under development at IT Research Center
( http://ciii.frc.utn.edu.ar ) .
×
(a)
(b)
Fig. 5. Consecutive images of the QA3 data set carpet, which are rotated 0 . 008rad
between them. The lack of intensity features make very dicult the use of traditional
feature trackers
Two mains goals were taken into account: maintain the accuracy of the se-
quential algorithm presented by Araguas et al. [10] and ensure a processing time
 
Search WWH ::




Custom Search