114
accuracy in complex environments. Nevertheless, we found that the accuracy of the results
decreases with the increasing complexity of the environment (e.g.
ecosystem and robot
environment). An interesting issue under investigation in subsequent and future works is
implementing/developing methods of high accuracy and efficiency for robot control in
very difficult environments.
115
Chapter 10
10.
Conclusion and Outlook
The Camera and the image processing unit belong the very important parts of machine
vision-based ADAS concepts. Different weather and environmental
conditions can
significantly influence the image processing performance. The key challenging issue in
ADAS
is safety; hence all part of the system should work under any conditions without
providing wrong information. To have a robust and real-time image processing module for
ADAS, we do need a high performance processing system.
Computing huge amount of
visual information for extracting meaningful data, features, etc needs a special/appropriate
hardware and processing architecture. In this research we did
a survey about different
hardware platforms for image processing and according to real-time related ADAS
requirements we have proposed a robust and real-time architecture based on CNN, FPGA
and GPU. The whole the system is a CNN based processing
which is developed
implemented on either GPU or FPGA. Before CNN implementation, we tried to understand
the theory of analog computing and developed a direct emulation of that paradigm on the
FPGA. Since there are many similarities between a CNN implementation and analog
computing
based on DDA, we could find a good way to implement CNN on FPGA. In this
thesis, we have shown that CNN can solve a series of image processing tasks in real-time.
This thesis has formulated 7 key research questions and each main chapter did provide an
answer to a respective research question:
