Ultra fast cnn based Hardware Computing Platform Concepts for adas visual Sensors and Evolutionary Mobile Robots

Download 3,22 Mb. Pdf ko'rish
bet	17/81
Sana	16.05.2024
Hajmi	3,22 Mb.
	#238917

1 ... 13 14 15 16 17 18 19 20 ... 81

Bog'liq
Alireza Fasih

2.3

Contribution for real-time ADAS
In this thesis, a new architecture for both hardware and software is proposed. In the
proposed platform concept we can perform the ADAS algorithms related to the visual
sensor processing within the fixed real-time constraint/deadline of 15 ms. Chapter 8
proposed a hardware and software model for implementing a robust ADAS system based
on GPU. As described with more details in another chapter of this thesis chapter 7; a
combination of FPGA and the “Cellular Neural Networks” paradigm does offer
a powerful
and robust platform concept that does ensure a real-time image processing for various
ADAS solutions. FPGA is a field reconfigurable hardware that is generally specified by a
hardware description languages such as VHDL, Verilog or SystemC [48]. The integration of
hardware (hard core) and software (soft core) within FPGA does represent a great
advantage of FPGA technology as it ensure a great flexibility and robustness. Due to the
huge amount of logic cells, basic logical operators and a routing system that is dedicated in
FPGA like a “switch matrix”, one is able to implement very complex functionalities for
processing data through FPGA.
Developers are able to realize the integration of FPGA and external I/O (i.e. input/output)
peripherals. Hence one can capture videos or data from the outside of the FPGA. In this
thesis, the integration of a HD-camera with the FPGA through a small daughter-board is
proposed. The FPGA logic cells and the internal CPU will have an access to the high-level
data through this daughter board and can load frames of data for processing. It is known
that FPGAs are perfectly suitable for performing parallel tasks for signal and image
processing. One of the promising paradigms for ultrafast image processing is based on
Cellular Neural Network
(CNN) [26]. Many forms/architectures/versions of CNN do exist,
that are discrete-time CNN (DT-CNN), non-linear CNN (NL-CNN), etc. But the DT-CNN
appears to be more appropriate for image processing because it does require less

22
hardware resources [25]. Each CNN cell consists of some basic mathematical operator such
as addition, subtraction, integration module, and a sigmoid function. Beside the integration
of CNN and FPGA as a target platform for ultrafast ADAS related image processing we do
also propose and have developed an alternative concept that does integrate CNN and
another also actually promising technology, the GPU. Details of this additional proposal
(that combines CNN and GPU) are presented in two other chapters of this thesis; see
chapters 7 and 8. The GPU technology does also offer a series of advantages ranging from
design flexibility, availability, costs, the possibility of an easy integration with other
framework such as
Open Computing Language
(OpenCL), and much more. The company
AMD has released an embedded GPU to provide high performance in mobile and
embedded systems. AMD Radeon E6760 is an embedded discrete graphics processor that
supports OpenCL and it has a good performance for parallel processing. Hence, having a
high-performance system in the scale of embedded system is possible.

23
Chapter 3

Download 3,22 Mb.

1 ... 13 14 15 16 17 18 19 20 ... 81

Download 3,22 Mb.

Pdf ko'rish