UiPath Test Automation using UiPath Test Suite series, part 2
Machine Vision for Intelligent Drones: An Overview
1. MACHINE VISION FOR INTELLIGENT
DRONES: AN OVERVIEW
APRIL 27TH 2016
KEVIN HEFFNER SAMUEL FOUCHER
RESEARCH ASSOCIATE, CRIM DIRECTEUR, VISION AND IMAGING TEAM
PEGASUS RESEARCH & TECHNOLOGIES CRIM
3. 3
noun: drone; plural noun: drones
a male bee in a colony of social bees, which does no work but can fertilize a queen.
Source: Steven Zaloga, Unmanned Aerial Vehicles: Robotic Air Warfare 1917-2007, Osprey Publishing, 2008
In 1935, U.S. Adm. William H. Standley saw a British demonstration of the Royal
Navy's DH 82B Queen Bee.
Commander D Fahrney was tasked to develop something similar for the US Navy.
"Fahrney adopted the name ‘drone’ in homage to the Queen Bee”
DRONE TERMINOLOGY
4. 4
TYPES OF DRONES BY SIZE
UAV ACRONYMS
Unmanned Aerial Vehicle (UAV)
Micro Aerial Vehicle (MAV)
Miniature Aerial Vehicle (MAV)
Small UAS (SUAS)
Medium Altitude Long Endurance (MALE)
High Altitude Long Endurance (HALE)
Remotely Piloted Aircraft Systems (RPAS)
9. 9
Which rules apply to me ?
http://www.tc.gc.ca/media/documents/ca-standards/Info_graphic_-
_Flying_an_umanned_aircraft_-_Find_out_if_you_need_permission_from_TC.pdf
START
NO
I have read and can meet
the exemption conditions
for UAVs < 2 kg
I have read and can meet
the exemption conditions
for UAVs from2 to 25 kg
You don’t
need permission,
but you must meet
the exemption
conditions
You must
apply for a
Special Flight
Operations
Certificate
You don’t
need permission, but
you do have to fly
safely
You don’t
need permission, but you
must meet the exemption
conditions AND notify
Transport Canada with:
1.Contact info
2.UAV model
3.Description of operation
4.Geographic boundaries
of operation
NO
It weighs > 35 Kg
NO
I use my aircraft for
WORK or RESEARCH
YES
YES NO
It weighs < 2 Kg
NO YES YESNO
YES YES
It weighs > 25 Kg
CANADIAN UAV REGULATION
10. 10
http://www.tc.gc.ca/eng/civilaviation/standards/standards-4179.html
• Fly your drone during daylight and in good weather (not in clouds or fog).
• Keep your drone in sight, where you can see it with your own eyes – not only
through an on-board camera, monitor or SmartPhone.
• Make sure your drone is safe for flight before take-off. Ask yourself, for
example, are the batteries fully charged? Is it too cold to fly?
• Know if you need to apply for a Special Flight Operations Certificate
• Respect the privacy of others – avoid flying over private property or taking
photos or videos without permission.
CANADIAN UAV REGULATION
• Closer than 9 km from any airport, heliport, or aerodrome.
• Higher than 90 metres above the ground.
• Closer than 150 metres from people, animals, buildings, structures, or vehicles.
• In populated areas or near large groups of people, including sporting events,
concerts, festivals, and firework shows.
• Near moving vehicles, highways, bridges, busy streets or anywhere you could
endanger or distract drivers.
• Within restricted and controlled airspace, including near or over military bases,
prisons, and forest fires.
• Anywhere you may interfere with first responders.
Do
Don’t fly
11. 11
http://www.tc.gc.ca/eng/civilaviation/standards/standards-4179.html
Currently there is no requirement for a Pilot or Operator License.
Transport Canada has communicated their intent to institute new
regulatory requirements that will address licensing and training of
drone operators for drones up to 25 kg.
These requirements also establish how the UAV should be marked
and registered.
CANADIAN UAV REGULATION
13. 13
DRONE EDUCATION
1. Make drones easier and safer to operate
2. To increase their autonomy and decrease human supervisory
control requirements
3. So that they can perform complex functions
4. So they can collaborate with other systems (including drones)
Why do we want drones to become more intelligent ?
Making decisions
Reasoning
Prioritizing tasks
Detecting differences
Finding similarities
Learning from experience
Characteristics of Intelligence
14. 14
*See Ingrand and Ghallab - http://homepages.laas.fr/felix/publis-pdf/aicom13.pdf
Cognitive Robotics – Breakdown of intelligence into functions
ROBOT INTELLIGENCE
Planning off-line prediction of feasible actions;
Acting task execution; refines planned actions;
Observing detects and recognizes;
Monitoring comparison of predicted versus observed;
Goal reasoning monitoring at mission level;
Learning acquire, adapt and improve through experience;
DELIBERATION FUNCTIONS*
20. 20
COMPUTER VISION
Object or place recognition
Visual odometry
3D reconstruction
Target tracking
Anomaly detection
Many applications
OpenCV
PCL
ArrayFire
…
Open-source libraries
The main goal of Computer Vision is to enable a computer to analyze, process
and understand one or more images taken by a vision system.
Development of MV Applications has exploded recently thanks to the advent of
high-res digital cameras along with increased computing power & machine learning.
21. 21
PLATFORM
Ubiquity of low-
cost prosumer
drones
SENSORS
Smarter, Smaller,
Integrated
Processors
PERFECT STORM OF TECHNOLOGIES
Size
Weight
and
Power
Price
(low) SWaPP
COMPUTING
RESOURCES
Embedded, cloud
Machine Vision
Libraries
22. 22
ON-BOARD COMPUTER VISION
The combination of computer vision and IMU data (altitude, acceleration, etc.)
can improve the precision and robustness of drone navigation.
Vision-based solutions are interesting for small drones (< 2.0 kg) for indoor
navigation, obstacle avoidance and other problems.
Navigation
Path Planning
Localization
Geo-fencing
Environment
GPS-denied
Obstacle avoidance
Swarming
Automatic Landing and takeoff
Toward increasing autonomy
Mission critical aspects
Environment mapping
Object detection
Anomaly detection
Video enhancement:
De-hazing
Image stabilization
Super-resolution
Contrast enhancement
Data compression
Application related processing
23. 23
SIMULTANEOUS LOCALIZATION AND MAPPING (SLAM)
Example of an application for an on-board vision system
Precise drone localization
Environment mapping around the drone
Sensing with a sonar, a laser or a camera
Object and place recognition
Problems to solve
Motion
Initial
position
prediction
Image
Keypoint
Extraction
Matching
Revised
Prediction
24. 24
KEYPOINTS DETECTION
Robustness
Appearance (illumination, occlusion, etc.)
Viewpoint
Local and distinct description
Quantity
Low computation cost
Ideal properties
Keypoints are sailient points in an image
Harris
SURF
SIFT
ORBE
FAST
Many different techniques
Detection
Description
Matching
25. 25
VISUAL SLAM
Camera + SLAM = monocular visual SLAM (vSLAM) exploits visual information
(keypoints)
RGB-D or stereo cameras directly minimize the photometric error
Matching between successive images Mapping
26. 26
VISUAL SLAM FOR GPS-DENIED ENVIRONMENTS
https://www.youtube.com/watch?v=VWWvjSHZCNo
High precision intertial navigation
units are costly and heavy
Laser range finders (costly and
heavy)
Cameras (visual odometry)
Possible solutions
Precision: +/- 2.5m
Not accurate in urban environment
Not possible to use indoors
GPS navigation limitations
28. 28
A LIST OF DRONE APPLICATIONS
Analysis of first 3136 exemption requests
authorized by the FAA for small drones (<25 Kg)
29. 29
NATURAL RESOURCE MANAGEMENT
RESEARCH APPLICATIONS
Source: Shahbazi, M., Théau, J. et P. Ménard (2014) Recent applications of unmanned aerial imagery in natural
resource management. GIScience & Remote Sensing 51
Based on review of 150 articles about the use of UAV imagery
30. 30
Improve water use
Pest management
Weed-identification
Irrigation and water distribution
Manned aircraft: ~$1,000/hour
Satellite Imagery: $$$
Multispectral sensors:
• Visible (R,G,B)
• Near-Infrared (NIR)
• Thermal Infrared (TIR)
• Resolution < 1 cm, > 12 Mpixels
Hyperspectral: 640 bands
LIDAR
AGRICULTURE APPLICATIONS
36. 36
HOTSPOT DETECTION USING OPTIMAL SEARCH PATTERN
Pegasus Research and Technologies Research Project
Rapid deployment
Low-cost lightweight IR sensors
Optimized search path
Applications for
For missing persons
Remnant forest fire detection
Security applications
HOTSPOT DETECTION
37. 37
OPEN-FIELD HOTSPOT DETECTION TRIAL
FOR LOCALIZED SEARCH AND RESCUE
1. Determine presence and approximate
location of candidate “missing person”.
2. High resolution camera is then pointed
at location and fed to SAR team.
3. System queries operator for next step
4. Future capability for automatic object
recognition.
HOTSPOT DETECTION
39. 39
ROOF INSPECTION APPLICATION
1. Automated data collection
a. 3D Model construction images
b. High-resolution texture images
2. 3D Model construction
3. Roof extraction and zone detection
4. High-resolution texture generation
5. Inspection annotations
6. Report generation
ROOF INSPECTION PIPELINE
51. 51
Personal Drones
TRENDS AND CHALLENGES: DRONE HYPE CYCLE ?
Agribotics
Cinema
Geomatics
PublicSafety
EnergySector
SiteInspection
52. 52
http://www.tc.gc.ca/eng/civilaviation/standards/standards-4179.html
Advances in remote sensing, embedded processing and cloud
computing technologies have created opportunities for new
applications.
The availability of low-cost prosumer drones adds another dimension
of opportunity.
Currently most drones are flying sensors.
Adding intelligence to drones can improve safety and allow for drones
to participate in data gathering and processing as part of an
information ecosystem, e.g. Internet of Things (IOT).
Regulation and safety aspects will be main factors in the adoption and
practical use of drones for commercial purposes.
CONCLUSIONS