|
Ultra fast cnn based Hardware Computing Platform Concepts for adas visual Sensors and Evolutionary Mobile RobotsBog'liq Alireza FasihList of Figures
Figure 1-1: ADAS processing levels Architecture
….…………………………………………………….....
..6
Figure
2-1: Comparison of two different ADAS system. (a) Lane departure warning, (b)
Licence plate recognition ................................................................................................................................. 20
Figure
3-
1: Shared memory’s parallel proce
ssing model ................................................................... 26
Figure
3-2: General idea for distributed processing ............................................................................. 27
Figure
4-1: Model of a feed-forward neural network with four inputs and one output ........ 30
Figure
4-2: Reshaping and normalizing 2D image data for loading in a MLP network .......... 34
Figure
4-3: Architecture of system for processing images based on CNN ................................... 36
Figure
4-4: CNN architecture based on GPU. ........................................................................................... 37
Figure
5-1: CNN Architecture ......................................................................................................................... 40
Figure
5-2: Basic architecture of CNN cell: the equivalent electrical circuit ............................... 44
Figure
5-3: Two images of a synthetically generated image sequence ......................................... 49
Figure
5-4: (a), (b), (c): Input image, edge extracted image and threshold image
respectively ............................................................................................................................................................ 50
Figure
5-5: (a) Input image and (b) thresholded image having no textured plane ................. 50
Figure
5-6: (a) initial condition image (b) input image (c) target image ..................................... 51
Figure
5-7: (a), (b): Two input images of a sequence; (c): Target Image; (d): CNN generated
output image. ......................................................................................................................................................... 52
Figure
5-8: Removing the rectangle part from the figure ................................................................... 53
Figure
5-9: Removing noise from the image ............................................................................................ 53
Figure
5-10: Filling a long tube with a dotted pixel .............................................................................. 54
Figure
5-11: Thresholding to a specified limit ........................................................................................ 55
Figure
5-12: Processing scenario by pre-calculated CNN templates ............................................. 56
Figure
5-13: Integration of CNN on FPGA with PowerPC for speeding up of genetic
algorithm ................................................................................................................................................................. 57
Figure
6-1: Flow diagram for modelling the Rössler Equation (see Equation (6-2)) in the
analog computing scheme emulated on FPGA ........................................................................................ 62
Figure
7-3: Xilinx XtremeDSP Kit 3400 ...................................................................................................... 73
Figure
7-4: Convolution and stream processing diagram .................................................................. 75
Figure
7-5: Scheme of the DDA based model for CNN .......................................................................... 78
Figure
7-6: Convolution diagram for the control template ................................................................ 78
Figure
7-7: Convolution diagram for feedback template .................................................................... 79
Figure
7-8: Impulse CoDeveloper Design Process Diagram .............................................................. 80
Figure
7-9: Real-time output of the system on monitor ...................................................................... 81
Figure
8-1: A simple CNN array architecture ........................................................................................... 88
Figure
8-2: System Design Architecture .................................................................................................... 90
Figure
8-3: A comparison traditional loop with the OpenCL data parallel kernel ................... 92
Figure
8-4: OpenCL Programming Flow .................................................................................................... 93
Figure
8-5: (a) Input image for CNN (b) Output image of CNN on GPU ........................................ 94
vii
Figure
8-6: Output of CNN on CPU after applying enhancement template ................................. 94
Figure
9-1: Shows the spatial wave and time domain on a CNN, which connected to the
robot actuators ..................................................................................................................................................... 97
Figure
9-2: System Architecture Diagram .............................................................................................. 100
Figure
9-3: Robot hinges connection to CNN array ............................................................................ 102
Figure
9-4: Template Encoding in an Array List .................................................................................. 102
Figure
9-5: Two-Point Crossover Method for
Template ‘A’
............................................................ 104
Figure
9-6: Snake robot lateral undulation locomotion ................................................................... 106
Figure
9-7: Wave generated for lateral undulation locomotion ................................................... 106
Figure
9-8: Wave generated for rectilinear locomotion ................................................................... 108
Figure
9-9: Snake Robot rectilinear locomotion .................................................................................. 108
Figure
9-10: Series 1 is fitness-function value; Series2 is fitness-function minimum value,
during cycle of time in learning process. (Series1 is error rate; Series2 is number of
itteration/time) ................................................................................................................................................. 109
Figure
9-11: Learning 4-legs semi-spider robot .................................................................................. 109
Figure
9-12: 4-leg robot spider, turning skill. ....................................................................................... 110
Figure
9-13: Moving 6-Leg Robot, around the Circle ......................................................................... 111
Figure
9-14: Wave generated for circular locomotion ...................................................................... 111
Figure
9-15: Broken Leg Spider as an unstructured robot ............................................................. 112
Figure
9-16: Wave generated for Broken-Leg Spider; Turning Skill ........................................... 113
|
| |
