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

 
7 
not trivial to speed up them by the using pipelining or parallel processing and it needs 
different model of processing [13]. 
1.2
 
Research questions and objectives of the thesis
In the following we do briefly describe and justify the key research questions of this thesis 
as well as the core of answers/solutions that have been obtained for each of them. Overall, 
we have formulated seven key questions concerning ADAS technology, importance of high 
performance computing and hardware implementation. 
 
Research question 1: What are the hard requirements of ADAS concerning real-time 
image processing and design flexibility? How far do traditional 
approaches fail to satisfy these requirements? 
One of the main issues to design an ADAS system is high processing speed and robustness 
of the environment perception (image processing based). The maximum interval for 
processing frames should not be greater than 20ms. In some cases such as lane departure 
warning and collision detection this level changes to maximum 10 ms processing time [14, 
15]. The fundamental issue in all image processing systems is robustness and accuracy. In 
ADAS technology which is based on image processing and machine vision we have to 
guarantee the stability and robustness of detection, identification and recognition of 
features [16]. For example a lane departure warning system should work in any 
environmental condition and lighting [16, 17].
Concerning video- based ADAS solutions different image processing filters and appropriate 
hardware architectures are required. Some filters are very complex and very demanding in 
terms of processing run time. Hence, we are interested in heavy parallel processing and 
task concurrency. In most of different ADAS concepts we one is using the same processing 
modules with different connectivity and topology. Hence, if we have a reconfigurable 
architecture or a universal model we do not need to reserve hardware resources for each 
different concept.
While following traditional methodological ways of doing and design one does use/involve 
sequential 
processing 
architecture, 
time 
sharing 
and 
multi-threading 

 
8 
algorithms/processing. The weakness of this traditional way of doing concerning 
processing performance is that it results in a ‘too long’ processing time and therefore
making it very difficult to satisfy real-time constraints. The real-time constraints expect a 
completion of the processing of a frame within a time window that is less than the 
capturing time of that frame. Therefore, for a 60 FPS (frames per second) rate we do have a 
maximum of 15 milliseconds to finish all the processing. And since one does generally need 
about 6~10 different high definition (HD) image preprocessing modules whereby each of 
them takes around 5ms per frame we thus do reach a total ranging between 30 ms and 
50ms of processing time if one tries to use the traditional processing schemes. This figure 
of 50 ms does fail to fulfill the real-
time constraint of “maximum 15 ms” processing time
per frame. It is therefore clear that traditional processing schemes are not capable of 
fulfilling the real-time constraints of visual sensors in/for ADAS.. Examples of ADAS 
solution of relevance for visual sensors are: the 
Lane Departure Warning
(LDW), 
Adaptive 
Cruise Control
(ACC), Emergency
 Brake Assistant
(EBA) and 
Blind Spot Detection
(BSD), etc.
The different ADAS solutions should be able to re-use the same components/modules for 
the image processing. Another weakness of classical algorithms they do not enable an easy 
re-use of functional components. Therefore we do need and are looking for a special 
architecture that is reconfigurable by software; such a concept will enable an easy re-use of 
the same platform for different functionalities and algorithms.

Download 3,22 Mb.