![](bg8.png)
5 CONCLUSION
One of the main questions here was whether a GPU
can perform the ball detection and flight path predic-
tion fast enough to achieve a high frame rate. The
results of the NVidia GTX 560 Ti GPU show exe-
cution times that enable data processing at 130FPS.
Moreover, using a GPU for the required calculations
proved to be a beneficial idea. The implemented pro-
gram was 3.46 to 7.17 faster when running on a GPU
than on a CPU; Intel i7-4770S. Using a newer GPU
would - most probably - accelerate the whole pro-
cedure. However, using this five years old GPU is
representative of GPUs in general. Additionally, as
earlier mentioned, integrated GPUs also offer compu-
tation possibilities. The performance we reached by
using the GPU shows that processing this flight path
prediction will most probably be possible also on an
embedded system with an onboard GPU.
We performed numerous optimizations to enable
such high-speed computation on the GPU used. Well-
thought-out algorithms and letting some data make a
detour over the fast shared memory were the keys to
success. However, some constraints still have to be
considered to achieve an execution time short enough
for this frame rate. Firstly, a small buffer is needed to
compensate for the maximum execution times, which
can be slightly too high for 110FPS (the two cam-
eras captured the scene with 110FPS). Secondly, the
Hough Circle Transformation turned out to be compu-
tationally too costly and time-consuming. Therefore,
the RANSAC algorithm had to be used to achieve the
desired execution time.
Additionally, the background subtraction step and
the adapting threshold profile used for the Canny
Edge Detector were necessary for achieving these ex-
ecution times. Otherwise, a considerably higher num-
ber of edge points would be in the images, which
would lead to an execution time that is too high for
110FPS. These additional computational steps also
enable a more precise detection. For the future, we
plan to consider an approach similar to (Tang et al.,
2015) to use more than one trajectory for the predic-
tion and examine the influences on the prediction ac-
curacy.
Detecting other - more complex - objects than a
ball will require both more computational power and
more convoluted algorithms. The detection will be
more complicated, and the prediction would have to
take also into account the object’s orientation with its
three degrees of freedom.
REFERENCES
Askari, M. et al. (2013). Parallel gpu implementation of
hough transform for circles. In IJCSI.
Barteit, D. et al. (2009). Measuring the intersection of a
thrown object with a vertical plane. In INDIN.
B
¨
auml, B. et al. (2011). Catching flying balls and preparing
coffee: Humanoid rollin’justin performs dynamic and
sensitive tasks. In ICRA.
Chen, S. et al. (2011). Accelerating the hough transform
with cuda on graphics processing units. Department
of Computer Science, Arkansas State University.
D’Orazio, T. et al. (2002). A ball detection algorithm for
real soccer image sequences. In ICPR.
Fung, J. et al. (2005). Openvidia: Parallel gpu computer
vision. In ACM Multimedia.
Gowanlock, M. et al. (2015). Indexing of spatiotemporal
trajectories for efficient distance threshold similarity
searches on the gpu. In IPDPS.
Hong, W. et al. (1995). Experiments in hand-eye coordina-
tion using active vision. In Lecture Notes in Control
and Information Sciences.
Jacobs, L. et al. (2013). Object tracking in noisy radar data:
Comparison of hough transform and ransac. In EIT.
Karimi, K. et al. (2010). A performance comparison of
CUDA and opencl. CoRR, abs/1005.2581.
Khronos, G. (2015). Opencl.
Luo, Y. et al. (2008). Canny edge detection on nvidia cuda.
In CVPRW.
NVidia, C. (2014). Cuda architecture.
NVidia, C. (2015). Geforce gtx 560 ti.
Ogawa, K. et al. (2010). Efficient canny edge detection us-
ing a gpu. In ICNC.
Pongratz, M. (2009). Object touchdown position prediction.
Master’s thesis, Vienna University of Technology.
Pongratz, M. (2016). Bio-inspired transport by throwing
system; an analysis of analytical and bio-inspired ap-
proaches. PhD thesis, TU Wien 2016.
Pongratz, M. et al. (2010). Transport by throwing - a bio-
inspired approach. In INDIN.
Pongratz, M. et al. (2012). Koros initiative: Automatized
throwing and catching for material transportation. In
Leveraging Applications of Formal Methods, Verifica-
tion, and Validation, pages 136–143.
Tang, X. et al. (2015). Efficient selection algorithm for fast
k-nn search on gpus. In IPDPS.
Wang, W. et al. (2011). Robust spatial matching for object
retrieval and its parallel implementation on gpu. IEEE
Transactions on Multimedia, 13(6).
Wu, S. et al. (2012). Parallelization research of circle detec-
tion based on hough transform. In IJCSI.
Parallelized Flight Path Prediction using a Graphics Processing Unit
393