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

Download 3,22 Mb. Pdf ko'rish
bet	13/81
Sana	16.05.2024
Hajmi	3,22 Mb.
	#238917

1 ... 9 10 11 12 13 14 15 16 ... 81

Bog'liq
Alireza Fasih

1.3.1 Scientiﬁc signiﬁcance of the thesis

1.3

Summary of the key contributions of the thesis
There are many different ADAS technologies offering various functionalities and levels of
safety. According to literatures, cameras are playing a very important role in modern ADAS
technology. Due to the large amount of visual data and the complexity of image processing
algorithms as well as of algorithms for extracting meaningful data (i.e. image enhancement,
noise removing, segmentation, classification, etc.) traditional methods fail to solve these
challenges and are difficult to implement in hardware. This thesis mainly targets the
development of concepts, hardware architecture(s) and related software concepts for
ultra-fast image processing in ADAS. From a scientific point of view the goal is designing a
fast and robust image processing system. CNN appears to be one of the best models that
offers sufficient potential to solve these challenges. Most of the advanced filters and
operators in image processing are mathematically based on PDEs. It is therefore
motivating that CNN is very appropriate for solving various types of PDE’s.
1.3.1

Scientiﬁc signiﬁcance of the thesis

The core objective in this thesis is the development of an ultra fast and robust computer
vision system based on cellular neural networks for ADAS. From a scientific perspective
the aim is to develop a platform and architecture for performing very complex image
processing filters and algorithms in dynamic environmental and different lighting
conditions. The proposed system is a CNN platform which is based on an emulation of
analog computing and can work in varying lighting conditions. This system opens the
doors for performing different PDE based image processing models on videos. The major
scientific contribution of this thesis is lying in the modeling of a CNN based image
processing architecture through an emulation of t
he traditional “analog computing” on
digital platforms (FPGA and GPU) to make a very robust and adaptive image processing
system. Flexibility of design is another main major factor for developers that have been

16
considered in this thesis. One can map most of the classical image processing models on
this system that then does operates in a parallel mode, and thereby ensuring high speed,
accuracy and robustness. This thesis did also focus on a template calculation concept
involving genetic algorithms. We have shown how to find appropriate templates for image
enhancement, obstacle detection and for controlling joints of an unstructured robot in
different scenarios.

Download 3,22 Mb.

1 ... 9 10 11 12 13 14 15 16 ... 81

Download 3,22 Mb.

Pdf ko'rish