83
Chapter 8
8.
Implementation of CNN on GPU
In this chapter the focus lies on the following research question:
“Can it be
demonstrated by a series of concrete image processing examples of relevance for ADAS, that
CNN based processing does really meet the hard requirements related to speed and
robustness?“
Every millisecond during driving is important. This could be a minimum interval for
processing signals in vehicles. In some cases such as Lane Departure Warning (LDW),
Adaptive Cruise Control (ACC), Emergency Brake Assist (EBA) and Blind Spot Detection
(BSD) system should take a decision in a few milliseconds. Main bottleneck for video based
ADAS is preprocessing image and preparation for extracting meaningful data from the
image. DSP board could be a good alternative solution but due to complexity of image
preprocessing, we do need a very robust and flexible platform which has potential of
cascading function, sharing memory, soft reconfiguration and high speed processing.
Hence, CNN architecture which implemented on FPGA or GPU could be more interesting
for us. In ADAS the sequence of image preprocessing is clear for designer; therefore a
flexible architecture which can reconfigure the system and change the functionality of the
module by a small matrix template would be very interesting for them. High definition
quality of image and performing complex filter in term of computation time, color space
conversion, and extracting features in ADAS technology are hard requirements related to
the speed and robustness.
|
