This document discusses augmented reality technology and visual tracking methods. It covers how humans perceive reality through their senses like sight, hearing, touch, etc. and how virtual reality systems use input and output devices. There are different types of visual tracking including marker-based tracking using artificial markers, markerless tracking using natural features, and simultaneous localization and mapping which builds a model of the environment while tracking. Common tracking technologies involve optical, magnetic, ultrasonic, and inertial sensors. Optical tracking in augmented reality uses computer vision techniques like feature detection and matching.
3. How do We Perceive Reality?
• We understand the world through
our senses:
• Sight, Hearing, Touch, Taste, Smell
(and others..)
• Two basic processes:
• Sensation – Gathering information
• Perception – Interpreting information
5. Reality vs. Virtual Reality
• In a VR system there are input and output devices
between human perception and action
6. Presence ..
“The subjective experience of being in one place or
environment even when physically situated in another”
Witmer, B. G., & Singer, M. J. (1998). Measuring presence in virtual environments: A presence
questionnaire. Presence: Teleoperators and virtual environments, 7(3), 225-240.
7. Slater, M., Banakou, D., Beacco, A., Gallego, J., Macia-Varela, F., & Oliva, R. (2022). A Separate Reality: An
Update on Place Illusion and Plausibility in Virtual Reality. Frontiers in Virtual Reality, 81.
Four Illusions of Presence (Slater 2022)
• Place Illusion: being in the place
• Plausibility Illusion: events are real
• Body Ownership: seeing your body in VR
• Copresence/Social Presence: other people are in VR
8. Senses
• How an organism obtains information for perception:
• Sensation part of Somatic Division of Peripheral Nervous System
• Integration and perception requires the Central Nervous System
• Five major senses (but there are more..):
• Sight (Opthalamoception)
• Hearing (Audioception)
• Taste (Gustaoception)
• Smell (Olfacaoception)
• Touch (Tactioception)
9. The Human Visual System
• Purpose is to convert visual input to signals in the brain
11. Sound Localization
• Humans have two ears
• localize sound in space
• Sound can be localized
using 3 coordinates
• Azimuth, elevation,
distance
12. Haptic Sensation
• Somatosensory System
• complex system of nerve cells that responds to changes to
the surface or internal state of the body
• Skin is the largest organ
• 1.3-1.7 square m in adults
• Tactile: Surface properties
• Receptors not evenly spread
• Most densely populated area is the tongue
• Kinesthetic: Muscles, Tendons, etc.
• Also known as proprioception
13. Proprioception/Kinaesthesia
• Proprioception (joint position sense)
• Awareness of movement and positions of body parts
• Due to nerve endings and Pacinian and Ruffini corpuscles at joints
• Enables us to touch nose with eyes closed
• Joints closer to body more accurately sensed
• Users know hand position accurate to 8cm without looking at them
• Kinaesthesia (joint movement sense)
• Sensing muscle contraction or stretching
• Cutaneous mechanoreceptors measuring skin stretching
• Helps with force sensation
14. Augmented RealityTechnology
• Combines Real and Virtual Images
• Needs: Display technology
• Interactive in real-time
• Needs: Input and interaction technology
• Registered in 3D
• Needs: Viewpoint tracking technology
15. AR Display Technologies
• Classification (Bimber/Raskar 2005)
• Head attached
• Head mounted display/projector
• Body attached
• Handheld display/projector
• Spatial
• Spatially aligned projector/monitor
16. Types of Head Mounted Displays
Occluded
See-thru
Multiplexed
23. AR RequiresTracking and Registration
• Registration
• Positioning virtual object wrt real world
• Fixing virtual object on real object when view is fixed
• Calibration
• Offline measurements
• Measure camera relative to head mounted display
• Tracking
• Continually locating the user’s viewpoint when view moving
• Position (x,y,z), Orientation (r,p,y)
25. Coordinate Systems
Local object
coordinates
Global world
coordinates
Eye
coordinates
Model transformation
• Track for moving objects,
if there are static objects as well
View transformation
• Track for moving objects,
if there are no static objects
• Track for moving observer
Perspective transformation
• Calibrate offline
• For both camera and display
27. The Registration Problem
• Virtual and Real content must stay properly aligned
• If not:
• Breaks the illusion that the two coexist
• Prevents acceptance of many serious applications
t = 0 seconds t = 0.5 second
28. Sources of Registration Errors
•Static errors
• Optical distortions (in HMD)
• Mechanical misalignments
• Tracker errors
• Incorrect viewing parameters
•Dynamic errors
• System delays (largest source of error)
• 1 ms delay = 1/3 mm registration error
29. Reducing Static Errors
•Distortion compensation
• For lens or display distortions
•Manual adjustments
• Have user manually alighn AR andVR content
•View-based or direct measurements
• Have user measure eye position
•Camera calibration (video AR)
• Measuring camera properties
32. Dynamic errors
• Total Delay = 50 + 2 + 33 + 17 = 102 ms
• 1 ms delay = 1/3 mm = 33mm error
Tracking Calculate
Viewpoint
Simulation
Render
Scene
Draw to
Display
x,y,z
r,p,y
Application Loop
20 Hz = 50ms 500 Hz = 2ms 30 Hz = 33ms 60 Hz = 17ms
33. Reducing dynamic errors (1)
•Reduce system lag
•Faster components/system modules
•Reduce apparent lag
•Image deflection
•Image warping
34. Reducing System Lag
Tracking Calculate
Viewpoint
Simulation
Render
Scene
Draw to
Display
x,y,z
r,p,y
Application Loop
Faster Tracker Faster CPU Faster GPU Faster Display
36. Reducing dynamic errors (2)
• Match video + graphics input streams (video AR)
• Delay video of real world to match system lag
• User doesn’t notice
• Predictive Tracking
• Inertial sensors helpful
Azuma / Bishop 1994
40. Frames of Reference
• Word-stabilized
• E.g., billboard or signpost
• Body-stabilized
• E.g., virtual tool-belt
• Screen-stabilized
• Heads-up display
41. Tracking Requirements
• Augmented Reality Information Display
• World Stabilized
• Body Stabilized
• Head Stabilized
Increasing Tracking
Requirements
Head Stabilized Body Stabilized World Stabilized
45. MagneticTracker
• Idea: coil generates current when moved in
magnetic field. Measuring current gives position
and orientation relative to magnetic source.
• ++: 6DOF, robust
• -- : wired, sensible to metal, noisy, expensive
Flock of Birds (Ascension)
46. InertialTracker
• Idea: measuring linear and angular orientation rates
(accelerometer/gyroscope)
• ++: no transmitter, cheap, small, high frequency, wireless
• -- : drifts over time, hysteresis effect, only 3DOF
IS300 (Intersense)
Wii Remote
47. UltrasonicTracker
• Idea: time of Flight or phase-Coherence SoundWaves
• ++: Small, Cheap
• -- : 3DOF, Line of Sight, Low resolution, Affected by
environmental conditons (pressure, temperature)
Ultrasonic
Logitech IS600
48. Global Positioning System (GPS)
• Created by US in 1978
• Currently 29 satellites
• Satellites send position + time
• GPS Receiver positioning
• 4 satellites need to be visible
• Differential time of arrival
• Triangulation
• Accuracy
• 5-30m+, blocked by weather, buildings etc.
49. Mobile Sensors
• Inertial compass
• Earth’s magnetic field
• Measures absolute orientation
• Accelerometers
• Measures acceleration about axis
• Used for tilt, relative rotation
• Can drift over time
51. Why Optical Tracking for AR?
• Many AR devices have cameras
• Mobile phone/tablet, Video see-through display
• Provides precise alignment between video and AR overlay
• Using features in video to generate pixel perfect alignment
• Real world has many visual features that can be tracked from
• Computer Vision well established discipline
• Over 40 years of research to draw on
• Old non real time algorithms can be run in real time on todays devices
52. Common AR Optical Tracking Types
• Marker Tracking
• Tracking known artificial markers/images
• e.g. ARToolKit square markers
• Markerless Tracking
• Tracking from known features in real world
• e.g. Vuforia image tracking
• Unprepared Tracking
• Tracking in unknown environment
• e.g. SLAM tracking
53. Visual Tracking Approaches
• Marker based tracking with artificial features
• Make a model before tracking
• Model based tracking with natural features
• Acquire a model before tracking
• Simultaneous localization and mapping
• Build a model while tracking it
54. Marker tracking
• Available for more than 10 years
• Several open source solutions exist
• ARToolKit, ARTag, ATK+, etc
• Fairly simple to implement
• Standard computer vision methods
• A rectangle provides 4 corner points
• Enough for pose estimation!
59. Marker Target Identification
• More targets or features à more easily confused
• Must be as unique as possible
• Square markers
• 2D barcodes with error correction
• E.g., 6x6=36 bits (2 orientation, 6-12 payload,
rest for error correction)
• Marker tapestries
59
60. But - You can’t cover world with ARToolKit Markers!
62. Natural Feature Tracking
• Use Natural Cues of Real Elements
• Edges
• Surface Texture
• Interest Points
• Model or Model-Free
• No visual pollution
Contours
Features Points
Surfaces
64. Natural Features
• Detect salient interest points in image
• Must be easily found
• Location in image should remain stable
when viewpoint changes
• Requires textured surfaces
• Alternative: can use edge features (less discriminative)
• Match interest points to tracking model database
• Database filled with results of 3D reconstruction
• Matching entire (sub-)images is too costly
• Typically interest points are compiled into “descriptors”
Tracking 64
Image: Gerhard Reitmayr
Image: Martin Hirzer
65. Tracking by Keypoint Detection
• This is what most trackers do…
• Targets are detected every frame
• Popular because
tracking and detection
are solved simultaneously
Keypoint detection
Descriptor creation
and matching
Outlier Removal
Pose estimation
and refinement
Camera Image
Pose
Recognition
66. What is a Keypoint?
• Invariant visual feature
• Different detectors possible
• For high performance use the FAST corner detector
• Apply FAST to all pixels of your image
• Obtain a set of keypoints for your image
• Describe the keypoints
Rosten, E., & Drummond, T. (2006, May). Machine learning for high-speed corner detection.
In European conference on computer vision (pp. 430-443). Springer Berlin Heidelberg.
69. Descriptors
• Describe the Keypoint features
• Can use SIFT
• Estimate the dominant keypoint
orientation using gradients
• Compensate for detected
orientation
• Describe the keypoints in terms
of the gradients surrounding it
Wagner D., Reitmayr G., Mulloni A., Drummond T., Schmalstieg D.,
Real-Time Detection and Tracking for Augmented Reality on Mobile Phones.
IEEE Transactions on Visualization and Computer Graphics, May/June, 2010
70. Detection and Tracking
Detection
Incremental
tracking
Tracking target
detected
Tracking target
lost
Tracking target
not detected
Incremental
tracking ok
Start
+ Recognize target type
+ Detect target
+ Initialize camera pose
+ Fast
+ Robust to blur, lighting
changes
+ Robust to tilt
Tracking and detection are complementary approaches.
After successful detection, the target is tracked incrementally.
If the target is lost, the detection is activated again
75. Marker vs.Natural FeatureTracking
• Marker tracking
• Usually requires no database to be stored
• Markers can be an eye-catcher
• Tracking is less demanding
• The environment must be instrumented
• Markers usually work only when fully in view
• Natural feature tracking
• A database of keypoints must be stored/downloaded
• Natural feature targets might catch the attention less
• Natural feature targets are potentially everywhere
• Natural feature targets work also if partially in view
76. Visual Tracking Approaches
• Marker based tracking with artificial features
• Make a model before tracking
• Model based tracking with natural features
• Acquire a model before tracking
• Simultaneous localization and mapping
• Build a model while tracking it
77. Model BasedTracking
• Tracking from 3D object shape
• Example: OpenTL - www.opentl.org
• General purpose library for model based visual tracking
82. Visual Tracking Approaches
• Marker based tracking with artificial features
• Make a model before tracking
• Model based tracking with natural features
• Acquire a model before tracking
• Simultaneous localization and mapping
• Build a model while tracking it
83. Tracking from an Unknown Environment
• What to do when you don’t know any features?
• Very important problem in mobile robotics - Where am I?
• SLAM
• Simultaneously Localize And Map the environment
• Goal: to recover both camera pose and map structure
while initially knowing neither.
• Mapping:
• Building a map of the environment which the robot is in
• Localisation:
• Navigating this environment using the map while keeping
track of the robot’s relative position and orientation
84. Parallel Tracking and Mapping
Tracking Mapping
New keyframes
Map updates
+ Estimate camera pose
+ For every frame
+ Extend map
+ Improve map
+ Slow updates rate
Parallel tracking and mapping uses two
concurrent threads, one for tracking and one for
mapping, which run at different speeds
85. Parallel Tracking and Mapping
Video stream
New frames
Map updates
Tracking Mapping
Tracked local pose
FAST SLOW
Simultaneous
localization and mapping
(SLAM)
in small workspaces
Klein/Drummond, U.
Cambridge
86. Visual SLAM
• Early SLAM systems (1986 - )
• Computer visions and sensors (e.g. IMU, laser, etc.)
• One of the most important algorithms in Robotics
• Visual SLAM
• Using cameras only, such as stereo view
• MonoSLAM (single camera) developed in 2007 (Davidson)
88. How SLAMWorks
• Three main steps
1. Tracking a set of points through successive camera frames
2. Using these tracks to triangulate their 3D position
3. Simultaneously use the estimated point locations to calculate
the camera pose which could have observed them
• By observing a sufficient number of points can solve for both
structure and motion (camera path and scene structure).
89. Evolution of SLAM Systems
• MonoSLAM (Davidson, 2007)
• Real time SLAM from single camera
• PTAM (Klein, 2009)
• First SLAM implementation on mobile phone
• FAB-MAP (Cummins, 2008)
• Probabilistic Localization and Mapping
• DTAM (Newcombe, 2011)
• 3D surface reconstruction from every pixel in image
• KinectFusion (Izadi, 2011)
• Realtime dense surface mapping and tracking using RGB-D
91. LSD-SLAM (Engel 2014)
• A novel, direct monocular SLAM technique
• Uses image intensities both for tracking and mapping.
• The camera is tracked using direct image alignment, while
• Geometry is estimated as semi-dense depth maps
• Supports very large-scale tracking
• Runs in real time on CPU and smartphone
93. Applications of SLAM Systems
• Many possible applications
• Augmented Reality camera tracking
• Mobile robot localisation
• Real world navigation aid
• 3D scene reconstruction
• 3D Object reconstruction
• Etc..
• Assumptions
• Camera moves through an unchanging scene
• So not suitable for person tracking, gesture recognition
• Both involve non-rigidly deforming objects and a non-static map
95. Combining Sensors andVision
• Sensors
• Produces noisy output (= jittering augmentations)
• Are not sufficiently accurate (= wrongly placed augmentations)
• Gives us first information on where we are in the world,
and what we are looking at
• Vision
• Is more accurate (= stable and correct augmentations)
• Requires choosing the correct keypoint database to track from
• Requires registering our local coordinate frame (online-
generated model) to the global one (world)
97. Types of Sensor Fusion
• Complementary
• Combining sensors with different degrees of freedom
• Sensors must be synchronized (or requires inter-/extrapolation)
• E.g., combine position-only and orientation-only sensor
• E.g., orthogonal 1D sensors in gyro or magnetometer are complementary
• Competitive
• Different sensor types measure the same degree of freedom
• Redundant sensor fusion
• Use worse sensor only if better sensor is unavailable
• E.g., GPS + pedometer
• Statistical sensor fusion
www.augmentedrealitybook.org Tracking 97
98. Example: Outdoor Hybrid Tracking
• Combines
• computer vision
• inertial gyroscope sensors
• Both correct for each other
• Inertial gyro
• provides frame to frame prediction of camera
orientation, fast sensing
• drifts over time
• Computer vision
• Natural feature tracking, corrects for gyro drift
• Slower, less accurate
99. Robust OutdoorTracking
• HybridTracking
• ComputerVision, GPS, inertial
• Going Out
• Reitmayr & Drummond (Univ. Cambridge)
Reitmayr, G., & Drummond, T. W. (2006). Going out: robust model-based tracking for outdoor augmented reaity. In Mixed and
Augmented Reality, 2006. ISMAR 2006. IEEE/ACM International Symposium on (pp. 109-118). IEEE.
102. ARKit – Visual Inertial Odometry
• Uses both computer vision + inertial sensing
• Tracking position twice
• Computer Vision – feature tracking, 2D plane tracking
• Inertial sensing – using the phone IMU
• Output combined via Kalman filter
• Determine which output is most accurate
• Pass pose to ARKit SDK
• Each system compliments the other
• Computer vision – needs visual features
• IMU - drifts over time, doesn’t need features
103. ARKit –Visual Inertial Odometry
• Slow camera
• Fast IMU
• If camera drops out IMU takes over
• Camera corrects IMU errors
105. Conclusions
• Tracking and Registration are key problems
• Registration error
• Measures against static error
• Measures against dynamic error
• AR typically requires multiple tracking technologies
• Computer vision most popular
• Research Areas:
• SLAM systems, Deformable models, Mobile outdoor tracking