Vladimir Kulyukin, Department of Computer Science, USU Vikas Reddy, Department of Computer Science, USU A Low Power Ad Hoc Computer Vision Testbed Network for EVs and Self-Driving Cars SELECT Annual Meeting and Technology Showcase – Logan, Utah – September 27-28, 2016 INTRODUCTION Computer vision (CV) can be used to optimize wireless energy transfers for EVs Wireless power transfer can be optimized by detecting magnetic charging stations and foreign objects on road surfaces However, advanced image processing algorithms may require expensive hardware and higher power, consumption so we wish to understand how CV algorithms can achieve higher accuracy and lower power consumption OPPORTUNITY : LEVERAGE POWER OF DISTRIBUTED COMPUTING CV accuracy increases with numbers of CPUs in that each CPU can be responsible for detecting specific features in complex road images However, higher CPU numbers require more power consumption, heat management, and packaging; they also increase communication overhead Leveraging an understanding of how CV accuracy is affected by numbers of nodes in ad hoc networks and power consumption requirements may lead to smaller, more robust vision-based packages that can be integrated into different EVs Lower power ad hoc networks can also be used as testbeds of different CV algorithms in different weather conditions and time periods FOREIGN OBJECT DETECTION WITH CONTOUR ANALYSIS • Crop the region of interest (ROI) with the station • Binarize and de-noise the ROI • Apply contour analysis to the ROI • Filter contours by pixel area CHARGING S TATION IDENTIFICATION WITH HOUGH TRANSFORM • Capture 360 x 240 frames from pi camera • Apply edge detector and probabilistic HT • Filter lines in ranges ± 45 ± 15 and ± 15 • Use topological line configurations for station identification EFFECTS OF GAUSSIAN BLUR (GB) LANE DETECTION WITH 1D HWT OBJECT DETECTION WITH 1D HWT CURVE DETECTION WITH 1D HWT SUMMARY & FUTURE WORK NODE COMMUNICATION: SFTP OVER WI-FI • GB eliminates noise and improves performance and detection accuracy in 720 x 480 frames • Tests indicate that it has little effect on accuracy in 360 x 240 frames A ROI is selection in the center of an image Edges are detected Rows and columns are selected 1D Ordered Haar Wavelet Transform (HWT) is applied to each column and each row Detected spikes signal presence or absence of objects of foreign objects on road surfaces Spikes identify smaller image regions where more sophisticated methods can be applied Horizontal segments are taken from left and right sides of captured frames 1D HWT is applied to each segment Detected spikes are used to identify presence of lanes Detected spikes identified in several consecutive segments can be connected into a line to identify a lane Same horizontal segments are taken from left and right sides of captured frames as in the case of lane detection 1D HWT is applied to each segment to detect spikes Our current work is focused on connecting spike centers in consecutive rows into curvatures Current ad hoc network consists of four Raspberry Pi computers; the network can function on a 13V battery for approximately 3 hours One daytime PiCam camera is connected to a Master node The ad hoc network has been tested on the USU EV bus and on a Jeep Wrangler Future work will focus on integrating on night vision Future work will also focus on improving curvature detection, foreign object identification, and power consumption requirements
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Vladimir Kulyukin, Department of Computer Science, USU
Vikas Reddy, Department of Computer Science, USU
A Low Power Ad Hoc Computer Vision Testbed Network for EVs and Self-Driving Cars
SELECT Annual Meeting and Technology Showcase – Logan, Utah – September 27-28, 2016
INTRODUCTION Computer vision (CV) can be used to optimize
wireless energy transfers for EVs
Wireless power transfer can be optimized by
detecting magnetic charging stations and foreign
objects on road surfaces
However, advanced image processing algorithms
may require expensive hardware and higher power,
consumption so we wish to understand how CV
algorithms can achieve higher accuracy and lower
power consumption
OPPORTUNITY: LEVERAGE POWER OF DISTRIBUTED COMPUTING CV accuracy increases with numbers of CPUs in that each CPU can be responsible for
detecting specific features in complex road images
However, higher CPU numbers require more power consumption, heat management,
and packaging; they also increase communication overhead
Leveraging an understanding of how CV accuracy is affected by numbers of nodes in
ad hoc networks and power consumption requirements may lead to smaller, more
robust vision-based packages that can be integrated into different EVs
Lower power ad hoc networks can also be used as testbeds of different CV algorithms
in different weather conditions and time periods
FOREIGN OBJECT DETECTION WITH CONTOUR ANALYSIS
• Crop the region of interest (ROI) with the station
• Binarize and de-noise the ROI
• Apply contour analysis to the ROI
• Filter contours by pixel area
CHARGING STATION IDENTIFICATION WITH HOUGH TRANSFORM
• Capture 360 x 240 frames from pi camera
• Apply edge detector and probabilistic HT
• Filter lines in ranges ± 45 ± 15 and ± 15
• Use topological line configurations for station identification
EFFECTS OF GAUSSIAN BLUR (GB) LANE DETECTION WITH 1D HWT OBJECT DETECTION WITH 1D HWT
CURVE DETECTION WITH 1D HWT SUMMARY & FUTURE WORK NODE COMMUNICATION: SFTP OVER WI-FI
• GB eliminates noise and improves
performance and detection accuracy in 720
x 480 frames
• Tests indicate that it has little effect on
accuracy in 360 x 240 frames
A ROI is selection in the center of
an image
Edges are detected
Rows and columns are selected
1D Ordered Haar Wavelet
Transform (HWT) is applied to
each column and each row
Detected spikes signal presence or
absence of objects of foreign
objects on road surfaces
Spikes identify smaller image
regions where more sophisticated
methods can be applied
Horizontal segments are taken from left and right
sides of captured frames
1D HWT is applied to each segment
Detected spikes are used to identify presence of
lanes
Detected spikes identified in several consecutive
segments can be connected into a line to identify a
lane
Same horizontal segments are taken from
left and right sides of captured frames as in
the case of lane detection
1D HWT is applied to each segment to
detect spikes
Our current work is focused on connecting
spike centers in consecutive rows into
curvatures
Current ad hoc network consists of four Raspberry
Pi computers; the network can function on a 13V
battery for approximately 3 hours
One daytime PiCam camera is connected to a
Master node
The ad hoc network has been tested on the USU EV
bus and on a Jeep Wrangler
Future work will focus on integrating on night
vision
Future work will also focus on improving curvature
detection, foreign object identification, and power
consumption requirements
Related Documents
